dsh-engram
English · 中文
Minimalist long-term memory for DeepSeek Harness, distilled from the
symbolic-index and pi-esr ideas —
with one goal: save tokens.
- Zero-LLM intake — auto-captures meaningful events from tool results by pure
pattern matching (git milestones with a written
-m commit message, edits to
key files, repeated errors), plus an explicit engram_store. Nothing on the hot
path calls a model, and pure plumbing — git push / git stash / silent
commits — is deliberately never recorded (see Auto-capture policy below).
Every write is run through a deterministic secret redactor first — API
keys, JWTs, Bearer tokens, private keys, AWS/Stripe/Slack/GitHub tokens and
key=value secret shapes are replaced with <REDACTED:…> markers before the
dedup hash, the char cap, and storage, so sensitive text can never land on
disk. Failing test runs are recognized by pure patterns (npm test,
node --test, vitest/pytest/jest, failure lines) and captured as
tags:["error","test"] entries with a raised signal. DSH goal-domain
integration: a completed or blocked goal (goal/change session event)
auto-sediments as a handoff/error memory (tag:goal), so the outcome of
a goal survives its session (see the GET /api/dsh-engram/goals read surface).
- User-correction / preference learning (
autoCapturePreferences, default on,
zero LLM) — the lesson that "when the AI makes a mistake and gets corrected,
that lesson should persist" (correx,
PAHF, Mem0 preference extraction): a
session/event listener watches user/message and turns deterministic
correction/preference cues into a dedicated preference memory kind —
Chinese (不要用X, 改用X, 以后统一用X…) and English (don't use X,
switch to X, prefer X, never use X…) patterns, three directions
(avoid / switch-to / prefer), with interrogative/one-shot-instruction/
memory-tool-meta-talk filters to keep false positives out. preference
memories outrank every other kind in the [ENGRAM] block ([R] prefix) so a
standing user rule is the first line the model sees each session; per-session
cap + recent-window dedup mirror the auto-capture discipline. Zero LLM, zero
embeddings, deterministic and pure — extracted clauses are bounded, and every
write still goes through the secret redactor.
- Symbolic index + progressive disclosure — a compact
[ENGRAM] block (default
budget 700 chars ≈ 175 tokens; one line per memory) is injected at prompt
assembly and frozen per session, keeping the request prefix byte-stable for
KV-cache reuse. The agent drills down with engram_recall / engram_detail instead
of dumping raw hits into context; recall ranks the in-domain pool with an
in-process BM25 pass (TF·IDF with label/phrase boosts plus a gentle time-decay
factor — deterministic, zero deps), appends a compact entity-neighborhood
table (reusing esr_link: node --rel--> node · conf%) for entity-anchored
hits, re-warms the old error entry when a new failure is near-identical
(repeat-failure revival: recency + hit climb toward promoteHits so the
failure resurfaces instead of piling up), and on zero local hits falls back
to DSH's own cross-session full-text index (ctx.sessionQuery, filtered by
cwd) instead of building a parallel SQLite index; when the FTS fallback also
comes back empty and the query contains CJK characters (Chinese FTS5
tokenization), recall runs a bounded substring scan over the newest session
logs (zstd-decompressed, LRU-cached) and appends matched past sessions —
deterministic # past sessions lines, capped files and bytes.
- Session-start auto-recall (
autoRecallOnStart, default on) — the pull-only
design had a gap proven in real use: recall tools relied on model initiative,
and in practice the model almost never calls engram_recall on its own (19
real sessions, thousands of tool calls, ~5 recall calls total). So at the FIRST
assembly of a new session the host runs one deterministic BM25 recall keyed on
the session's first user message and injects the top ≤ autoRecallLimit hits
as a [RECALL] block (default 700-char budget, "few and precise" — agentmemory
candidate ①). Relevant memory is in context even without a single recall call.
Pure rule, zero LLM, pure read (never bumps hits); superseded stale truth does
not occupy an injection slot (still fetchable via engram_detail); frozen per
session together with [ENGRAM] for prefix stability.
- Read-path activation (D) — a hit whose one-line preview is truncated
(>140 chars) carries a
… marker and the block's trailing guidance names that
concrete id: #<id> truncated — engram_detail <id> (full) · engram_recall <query>.
So the model knows the full record is one deterministic engram_detail call
away instead of assuming the preview is the whole memory (proven gap in real
use: 66 sessions, zero engram_detail calls — the detail path was invisible).
- Memory-to-memory semantics —
engram_store accepts optional
supersedes / contradicts memory ids (validated against the same
workspace). A superseded ("stale truth") memory is demoted to the tail of
recall and excluded from the [ENGRAM] block, while the superseding
statement keeps the line; a contradicted memory stays ranked but is flagged
· contradicted by <id>. Old rows are never deleted — re-fetchable, just
honestly ranked. Opt-in autoSupersede config (default off) auto-marks
entity-anchored replacement updates ("改用 / no longer / switched…") as
superseding their matching older row; explicit supersedes always wins.
- Failure→fix closure (zero LLM) — error memories carry a
cmd: signature
tag; when that same command later succeeds, engram auto-sediments a
procedure memory (fixed: <cmd> — N earlier failing runs now succeed),
tags the earlier error rows resolved (never re-ranking them), and recall
then surfaces the fix instead of the stale failure.
- ESR-lite closure protocol —
esr_task / esr_close / esr_link give tasks
a draft → active → stable lifecycle where stable requires real evidence
(artifact / evaluation / memory_ref), surfacing closure gaps instead of
letting the agent declare victory without proof. With verifyArtifact on
(default), a non-URL artifact is resolved against the workspace (= the session
cwd) and must exist on disk — otherwise the task stays ACTIVE with the reason;
force:true (or disabling the toggle) skips the disk check, the three gates
are still mandatory. The tool and the web forms share one gate
(store.evidenceGate), so the two surfaces can never drift.
Tasks are never isolated graph nodes: esr_task(entity=…) one-call wiring
auto-creates the ent_<slug> node and hangs the task on it
(task --relates_to--> entity, idempotent), so a task created with a domain
anchor immediately shows a relation; and esr_dep writes the dependency edge
into both the task's deps (blocker logic) and the shared links table, making
every task dependency a first-class visible edge in the graph.
- ESR trigger mechanism (pi-esr-aligned, hybrid: static protocol + frozen
snapshot + monotonically accumulated actionables + pull) — prefix-cache
stability is the point. Teaching the model WHEN to use ESR is a static
methodology section (byte-identical every turn); the
[ESR] block is a
deterministic per-session frozen snapshot (set at session start,
deterministically sorted, timestamps excluded, explicitly "WILL NOT
auto-refresh") plus the session's monotonically accumulated actionables:
the promote: / root-cause: / close: / stale: / escalate: lines are
APPENDED exactly once, the moment each matures, and stay frozen for the rest
of the session (never removed) — so the prefix changes only when genuinely
new decision-point info appears, never per-turn, while the timely nudges
(narrow funnel / recurring failure / work done / stale / balance) stay in
context. Live state is still pulled via esr_status (full deterministic
view + since_revision short "unchanged" response + derived actionables).
The tools/result recorder feeds the actionables' data sources; P4
conversion metering stays (#suggest-* tags + 10-minute attribution +
GET /api/dsh-engram/triggerstats). Pure rules, zero LLM.
- Session-end todo auto-sink (draft safety net,
autoSinkTodosOnEnd default
on) — during a session the native todo stays cheap, per-session working
memory that dies with the session. But when a session closes at its teardown
edge (session/disposed) while its plan still has pending todos, they are
auto-landed as ESR draft tasks (name = the todo text, description
marked 「源自会话计划」, deduped against existing task names, capped by
maxTasksPerWorkspace), so the plan never silently evaporates. Sinking to
draft — not active — keeps them out of the [ESR] active rows and creates
no evidence obligation; an explicit esr_claim / esr_task step is still
what promotes a draft to active. The "nothing auto-written during the
session" trade-off is unchanged; only the end-of-session data-loss gap is
closed. Disable via profile patch or the settings card.
- Memory GC (pi-esr constraints) — a scheduled, mechanical, archive-only
sweep: TTL-expired memories are archived, over-cap workspaces evict the
lowest-value entries, stable tasks past their retention window leave the
[ESR] surface, and dangling link edges are dropped. The working set (active
task refs, task memories, indexed hits) is never touched, and nothing is
hard-deleted — every archived entry keeps its id and stays re-fetchable.
- Context GC (auto-GC = replaces DSH's auto-compact) — NOT the memory-panel
GC. It takes over the host's
compaction service and swaps DSH's lossy
LLM-summary compression for mechanical eviction + re-fetch pointers:
evicted turns are scanned for engram/ESR anchors (#memory-id, tsk_*,
ent_*, file_path) and replaced with one line "this detail lives at
engram_detail / engram_recall / the [ESR] block"; only un-provenanced
turns get a scoped LLM narrative (gcNarrative, default on, off = fully
mechanical). All six GC constraints hold (working-set-protected,
pointer-salience, no-provenance-no-evict…); any error falls back to the
default compressor — compaction is never broken.
- Web viewer — a memory browser with benchmark-ish stats and a config card,
built entirely on DSH's native settings slots (no third-party UI package).
MIT · node >= 22.19 · host-half + browser-half in one package
Why another memory plugin?
Surveys of the existing DSH plugin ecosystem show the recall-bridge, approval-gate,
LLM-distillation and vector/graph niches are already crowded. dsh-engram fills the
three gaps that matter for token discipline:
- No model in the write path — capture is deterministic pattern matching.
- No raw text in the prompt — a bounded symbolic index is injected, plus the
top ≤ N memories auto-recalled for the session's first message (within budget);
deeper retrieval stays on demand ("retrieved ≠ full-text injected").
- Honest task closure — STABLE cannot be declared without evidence.
DSH already provides cross-session FTS (ctx.sessionQuery), storage
(ctx.storageDomain), prompt-injection hooks and settings slots; dsh-engram is a
thin composition layer over them, not a re-implementation.
Security model
The DSH plugin ecosystem is young and unvetted — there is no official directory,
no signature check, and the harness's read/write permission tiers do not
constrain what a plugin's own code can do. So the trust bar for a memory plugin
is higher than for a tool: a memory plugin sits on your prompt prefix and writes
to disk on your behalf. dsh-engram's model, stated plainly:
- No network, ever. The plugin has no outbound socket, no telemetry, no
update pinger. Nothing here can leak your sessions off the machine.
- No shell, no filesystem access beyond one storage file. It never executes
commands and never touches arbitrary paths. All data lives in the single
storage-domain unit
~/.dsh/storages/dsh_engram.json (plus the session-log
reads it makes through DSH's own ctx.sessionQuery).
- Secrets are redacted before they can land on disk. Every write path —
engram_store, auto-capture, and the import/restore route — runs text through
the deterministic redactText rule engine (API keys, JWTs, Bearer tokens,
private keys, key=value secret shapes) before the dedup hash, the char
cap, and storage. Nothing sensitive is ever persisted, so nothing sensitive
can be recalled later.
- Doing harm requires your harness's own permissions. The plugin only asks
the host to do things via the same seams the tools you already trust use. It
does not widen sandbox policy or escalate to
danger-full-access — if you
keep your profile's defaults, engram inherits exactly those bounds.
- Read-mostly web API, loopback-fenced. The GUI routes under
/api/dsh-engram reject any non-loopback caller unless you explicitly
whitelist a hostname via trustedHosts; even then the same-origin fence holds.
- Verifiable.
npm run dsh-engram -- doctor (/scripts/dsh-engram.mjs doctor) reports what the plugin is configured to touch; the whole store is
inspectable, nothing is opaque. If you ever doubt a claim in this section, the
code to check is lib/redact.js, lib/store.js and this plugin's
cordis.patch.yml.
What this deliberately does not claim: it is not an air-gap. If your profile
grants the agent danger-full-access or a shell, the agent can use those —
engram is a memory layer, not a sandbox. The point above is that dsh-engram
itself adds no surface of its own.
Data contract: export / import
Memory should not be a hostage. The four tables (memories / tasks / links /
entities) can be dumped and restored through a stable, versioned, machine-readable
shape — so the corpus survives a reinstall, a machine move, or (when a thin
adapter is written) a move to another harness, MCP end-point or Claude Code
skill. The contract is plain JSON with a self-describing header:
GET /api/dsh-engram/export?workspace=/path/to/project # one workspace
GET /api/dsh-engram/export # every workspace
{
"meta": { "format": "dsh-engram/export", "version": 1, "exportedAt": 1724…, "workspace": "/path/to/project" },
"memories": [ /* all rows, archived included — a backup never loses provenance */ ],
"tasks": [ /* draft/active/stable + archived */ ],
"links": [ /* typed edges */ ],
"entities": [ /* graph nodes */ ]
}
Restore is idempotent by default and gated when destructive:
POST /api/dsh-engram/import # body: { payload, mode, dryRun?, workspace?, confirm? }
mode: "merge" (default) — writes only rows whose id is absent, so a
backup can be re-applied any number of times without duplicating anything. Each
row keeps its own workspace.mode: "replace" — wipes the target workspace from all four tables, then
restores only its rows. Destructive, so it additionally requires
confirm: "restore" in the body.dryRun: true — computes the identical plan (what would be written /
skipped) without writing a single row.- Import honors the same invariants as
storeMemory: memory text passes the
secret redactor before touching disk, and rows that would break the contract
(missing id/workspace, empty text, over maxMemoryChars, id already present
in merge, over the workspace memory cap) are skipped and listed in the
response report instead of aborting the batch.
A restore round-trip for one workspace is then:
curl -s "http://127.0.0.1:3080/api/dsh-engram/export?workspace=$PWD" -o engram-backup.json
curl -s -X POST http://127.0.0.1:3080/api/dsh-engram/import \
-H "content-type: application/json" \
-d "{\"payload\": $(cat engram-backup.json), \"mode\": \"merge\"}"
Compatibility
- DSH 版本 / DSH versions:
>=0.1.2-alpha.2 <0.3.0-0 (本版 0.9.0; 已发布的 0.3.7、0.4.0 仍声明旧的 0.1.x 范围 >=0.1.2-alpha.2 <0.2.0-0 — 见下方矩阵)
- 最后验证日期 / Last verified: 2026-10-27 — DSH
0.2.1-rc.2 (dsh --version), web profile live composition; npm test 483/483
- 安装方式已验证 / Verified install:
dsh plugin --profile web add dsh-engram
Short version. Every published release (0.2.0 … 0.3.6) supports
the old DSH API graph (0.1.0-rc.7 / 0.1.1, peers ^0.1.0-rc.7)
only. DSH 0.1.2-alpha.2 and newer are not compatible with those
releases — the settings/client API family broke there. New-DSH support
starts with 0.3.7, the first release of the new API generation
(described below).
| dsh-engram release | Status | DSH support | Declared via |
|---|
0.2.0 … 0.3.6 (npm, GitHub v0.3.3–v0.3.6) | published | old DSH only — 0.1.0-rc.7 / 0.1.0-rc.8 / 0.1.1-rc.1 / 0.1.1-rc.2 | peers @deepseek-ai/dsh-*@^0.1.0-rc.7 |
0.3.7 / 0.4.0 (npm) | published | new DSH, settings installSection — >=0.1.2-alpha.2 <0.2.0-0 | engines.dsh + peers >=0.1.2-alpha.2 <0.2.0-0 |
0.5.0 (npm) | published | DSH 0.1.2-alpha.2 … <0.3.0-0 — 设置卡迁移到 entry-id volatile Config（0.1.7+/0.2.x），≤0.1.5 自动回退 installSection | engines.dsh + peers >=0.1.2-alpha.2 <0.3.0-0 |
0.5.1 | published | 同 0.5.0 — 会话扫描修复（session.v<N>.jsonl.zstd，issue #8）、数据导出/导入 GUI、替换/冲突可视化（issue #3） | 同 0.5.0 |
0.6.0 | published | 同 0.5.0 — Context Ledger、BoundaryPrune/ResultPack 默认开 + 账本否决、JIT 推送、Context GC 改 opt-in | 同 0.5.0 |
0.7.0 | published | 同 0.5.0 — session episode、resume brief、entity pages/cards、revalidatable anchors + layered TTL | 同 0.5.0 |
0.7.1 (npm) | published | 同 0.5.0 — web-plane Context GC on DSH 0.2.0 declarative presets (14-space rows), bundle patch re-enables the host tool-result-pruner, compaction savings feed the Context Ledger | 同 0.5.0 |
0.7.2 (npm) | published | 同 0.5.0 — L1 boundary prune arms through ctx.inject, fixing the boot-ordering race where a pruner activating one tick after the plugin stayed dormant for the whole process | 同 0.5.0 |
Where the API family broke — 0.1.2-alpha.2
0.1.2-alpha.2 is the first version of the new API family and the first that
breaks the released plugin versions:
@deepseek-ai/dsh-settings removed settingsNamespace() and
installSettingsSection() — the exact imports every release ≤ 0.3.6 uses —
and replaced them with settings.installSection. The old host half fails to
load on 0.1.2-alpha.2+.- The web client half swapped its inject set:
@deepseek-ai/dsh-client-runtime
no longer exists in the new family (replaced by dsh-api-remotes plus
dsh-client-locale / -conversation / -renderer / -ui-settings-plugins),
and the settings card used to talk through ctx.remote.settings.
- Supporting bumps at the same boundary:
@deepseek-ai/cordis ^4.0.1 → ^4.0.2,
@deepseek-ai/schemastery ^3.18.1 → ^3.18.2, and @deepseek-ai/dsh stopped
shipping config/agent-presets in the package (moved to dsh.configTrees).
The web config card no longer depends on any of this: since 0.3.7 it reads
and writes the dsh-engram settings namespace through the plugin's own
HTTP route (/api/dsh-engram/settings, same-origin fetch like the memory
viewer), so the card is generation-agnostic on both sides of the
0.1.2-alpha.2 boundary.
One more settings break, at 0.1.7 / 0.2.0 — and the reason 0.4.0 was
not actually loadable on the current DSH: @deepseek-ai/dsh-settings removed
settings.installSection again in 0.1.7-alpha.2 (and keeps it removed in
0.2.x). The new settings model is entry-keyed: editable fields are
.volatile() fields of the plugin's own Config (validated + hot-committed by
the loader), and the settings form is keyed by the profile entry id (engram),
so nothing needs "mounting". 0.5.0 feature-detects the two generations in
installEngramSettings:
- DSH
≥0.1.7 / 0.2.x — the 34 GUI knobs are the volatile Config fields
(lib/settings.js makeEngineConfigSchema, exported as Config from
lib/index.js); the card reads/writes entry engram through the same
/api/dsh-engram/settings HTTP route, and the service.settings seam
re-syncs resolved on loader/volatile-update so card edits land live on
new sessions.
- DSH
≤0.1.5 — the legacy settings.installSection namespace (dsh-engram)
is registered as before; nothing changes.
The declared range >=0.1.2-alpha.2 <0.3.0-0 covers both, and the HTTP wire
namespace stays dsh-engram on both generations (the seam's active namespace
drives mutations).
So: keep DSH on 0.1.1-rc.2 or older while running a release ≤ 0.3.6, and
upgrade the plugin to >=0.3.7 before moving DSH to 0.1.2-alpha.2.
Current release — 0.7.2
L1 boundary prune was silently dormant on some profiles — a boot-ordering
race, not a missing patch.
BoundaryPrune resolved DSH's toolResultPruner with a strict
ctx.get("toolResultPruner") at install time. On DSH 0.2.0 web profiles the
pruner's fiber activates ~100 ms after the plugin (measured on the live
composition dsh-context + dshmarket + dsh-engram: t0 strict=undefined / loose=object, strict resolves 100 ms later), so the strict lookup was empty,
installBoundaryPrune gave up, and L1 never armed for the life of the process —
taking the ledger's only prune-credit leg with it. The status snapshot also
reported that race as "row missing", pointing at the wrong fix.
- Arm through
ctx.inject(["toolResultPruner"], …) — cordis runs the
callback the moment the service becomes active (and again after a reload), so
arming no longer depends on composition order; the durable status snapshot is
rewritten as soon as L1 is armed.
- "Absent" and "late" now read differently —
boundaryPruneAvailability
falls back to a non-strict lookup: a disabled/missing row still says
no toolResultPruner service … (the bundle patch must re-enable …), a
registered-but-inactive one says registered but its fiber is not active yet.
- Verified — the live bundle composition (with
dsh-context) reproduces
armed:false in an isolated DSH_HOME; after the fix the same composition
reports armed:true · threshold 8192. npm test 406/406.
Previous release — 0.7.1
Web-plane Context GC on DSH 0.2.0's declarative presets — the preset
format changed shape (one @deepseek-ai/dsh-agent-preset declaration inside
dsh-web-app/presets/*.patch.yml), and the auto-wiring was anchored to the old
agent.cordis.yml layout, so autoWebCompaction silently skipped every preset
(and could half-rewrite a stock row: the pattern did not include its trailing
newline).
- Structural row location — the
compaction group is found by YAML
structure (a list item carrying group: true), not by column position; the
stock/engram rows are validated by BOTH their - id: and name: values, and
the replacement inherits the row's own indent. Legacy agent.cordis.yml
(4 spaces) and declarative patch files (14/16/18 spaces) both round-trip
byte-identically.
- Declarative discovery — the preset roster no longer carries
path on
DSH 0.2.0; the plugin maps each roster id to the shipped
<presetsDir>/<id>.patch.yml (resolved through the dsh install, with
DSH_WEB_PRESETS_DIR override) and skips user-defined declarative presets
(profile-patch overrides) with an explicit log line instead of failing.
- Status/doctor —
npx dsh-engram status/doctor discover both preset
generations, show the detected layout, and flag autoWebCompaction=false
(the opt-in toggle that makes the web plane a no-op), node_modules preset
wiring (survives npm update dsh only while auto-wiring stays on), and
profile-patch preset-* overrides that outrank the shipped files.
- Migration — restart
dsh web with autoWebCompaction:true: the next boot
re-wires every stock declarative preset (idempotent refresh; existing
hand-wired rows are rewritten to the requested engine config). npm test
396/396.
Previous release — 0.7.0
From event records to maintainable current truth: sessions sediment an
episode at teardown, new sessions get a resume brief, esr_node becomes a
readable page, and file anchors make stale truth detectable.
- Session episode — zero-LLM digest at
session/disposed (files / commands /
failures→fixed / decisions / open loops), never indexed, newest 20 per
workspace kept.
- Resume brief — the newest episode from the last 7 days is injected as a
[RESUME] line (≤240 chars + drill pointer) at session start.
- Entity pages + cards —
truth / timeline / open loops / conflicts / relations; up to 2 one-line cards in the index, full page via
engram_detail(entity=…).
- Revalidatable anchors — file-backed memories record
mtime/size/sha256;
a bounded revalidation at session start tags changed/missing sources stale
(dropped from index/truth, still re-fetchable).
- Layered TTL — preference/procedure permanent, episode 14d, error 60d,
others 180d.
- Migration — only new writes are anchored (no historical backfill);
restart
dsh web after upgrading; an episode appears after one session ends,
the resume brief on the next new session. The 0.6.0 ledger / defaults / JIT
behaviour is unchanged.
The published 0.4.0 declared >=0.1.2-alpha.2 <0.2.0-0 — which the semver
gate rejects on every DSH 0.2.0-0 prerelease (<0.2.0-0 excludes
0.2.0-rc.x), and its host settings half still called the installSection
API that dsh-settings 0.1.7+ removed. 0.5.0 fixes both: the range
opens to <0.3.0-0 (verified against DSH's own includePrerelease check, so
0.2.0-rc.2 and the whole 0.2.x line pass), and the settings card is
migrated onto the entry-id volatile Config model described above (with a
feature-detected legacy fallback).
| Component | Supported range |
|---|
| DSH runtime / host / client package family | >=0.1.2-alpha.2 <0.3.0-0 |
@deepseek-ai/cordis | ^4.0.2 |
@deepseek-ai/schemastery | ^3.18.2 |
| Node.js | >=22.19.0 |
0.1.2-alpha.2 is the minimum version for this API family: host settings are
feature-detected — on DSH ≤0.1.5 they register the dsh-engram
namespace via settings.installSection; on ≥0.1.7 / 0.2.x the form is the
plugin's Config with the 34 GUI knobs marked .volatile() (entry id
engram). The browser settings card reads and writes through the plugin's own
HTTP route (/api/dsh-engram/settings, see Route B above) on both. The full
regression suite (npm test 396/396, incl. the new settings-v2 coverage for the
volatile-Config model) passes against the DSH 0.2.0-rc.2 dependency set
(cordis 4.0.4 + schemastery 3.18.4 + cosmokit, matching the global DSH CLI).
The .volatile()/loader hot-commit path and the describe-side projection were
verified against the real dsh-settings 0.2.0-rc.2 utilities, and a full
real-machine boot smoke on an isolated dsh 0.2.0-rc.2 profile passed:
the plugin boots with no version-range denial, the settings card form is
served for entry engram (all 27 GUI knobs, incl. trustedHosts), and a
GET → PUT (mutate ops, revision 0→1) → unset round trip on
/api/dsh-engram/settings lands in the profile row and reads back. The older
0.1.0-rc.7 and 0.1.1 API graphs are outside this release's compatibility
range.
To install the DSH version used for validation:
npm install --global @deepseek-ai/dsh@0.2.0-rc.2
dsh --version
Install
# from GitHub (this repo)
dsh plugin --profile web add github:skepsun/dsh-engram
# once published to npm
dsh plugin --profile web add dsh-engram
# local development (symlink — edits apply immediately)
dsh plugin --profile web add link:/path/to/dsh-engram
Then restart dsh web. Data persists in ~/.dsh/storages/dsh_engram.json.
The npm and GitHub installs below need no manual dependency step: pnpm
installs zod and vendors the optional @deepseek-ai/* peers into the
plugin's own node_modules, and the CLI auto-registers the plugin into the
profile's dsh.profile.bundles. The setup-links step below is only for
the link: development workflow, where pnpm deliberately does not install a
symlinked directory's dependencies.
A fresh session is required to see the injected [ENGRAM]/[ESR] blocks and the
tools; both prompts and the tools registry are assembled per session.
Dependencies for link: installs
A symlinked plugin resolves its imports from its own node_modules, so the
host-side dependencies must be present next to the checkout — they are not
tracked by git:
cd /path/to/dsh-engram
node scripts/setup-links.mjs # one command: links the @deepseek-ai
# workspace packages into node_modules AND
# installs zod (reused from the harness
# pnpm store, or via `npm install`)
The script auto-locates the harness checkout at ../deepseek-harness (also
works when it sits next to the repo's parent, e.g. E:\deepseek-harness +
E:\kototoro_demo\dsh-engram); override with DSH_HARNESS_DIR. Without this
step, dsh web boot fails with ERR_MODULE_NOT_FOUND: Cannot find package 'zod' (and would fail on the @deepseek-ai/* peers next). node scripts/setup-links.mjs --check prints the state without writing anything.
What you get in the GUI
After restart, inside the native DSH settings surface:
-
Sidebar "ESR 看板" entry + full-screen kanban — one more row under
New Session with a live active-task badge (polled from /overview every
30s across all workspaces). Clicking it opens a full-screen board in the
center column: 草稿 / 进行中(gaps) / 就绪(evidence ready) / 已闭环 columns,
workspace filter + search, an inline create form, and per-card
"补齐证据 → 关闭" closure forms sharing the esr_close gates (artifact +
evaluation + memory_refs). A 看板 / 图谱 (board/graph) toggle in the
header reuses the full relation graph (esr_node/esr_link force-directed SVG:
entity circles + task check badges, drag/zoom/select for relation details),
following the workspace filter and refreshed by the same 20s polling.
Following the task-board precedent, the entry and
the board are DOM-mounted and self-heal (MutationObserver re-inserts on shell
re-renders), with cross-panel exclusivity against task-board/ssh (opening one
evicts the others; clicking a sidebar session/workspace row hands the center
column back to the conversation). The conversation subtree stays mounted
underneath and is hidden by html[data-dsh-engram-board-active], so toggling loses
no state.
-
Unified task strip above the composer — the conversation dock that DSH
ships for its built-in todo tool is taken over (same conversation.input.dock
cell / id: todo at a lower priority) and merged into one modern control:
the session's current plan (todo_write's todos projection) plus the
workspace's persistent ESR tasks (with evidence-gap badges and an inline
"补齐证据 → close" form) plus the relation graph rendered as
node → relation → node chips with entity/task names resolved. It only shows
while there is something to show, stays live with 15s polling, and the built-in
plan still renders (without the ESR parts) if the loopback-fenced API is
unreachable. A workspace-switcher chip leads the strip: it defaults to
following the current session (the tooltip says so), and its dropdown pins the
ESR task/relation source to any workspace (✓ marks the active pin; the × or
"follow session" entry reverts). Switching refetches immediately and is a pure
UI focus change — the model's session context and the per-session frozen
injection blocks are untouched.
-
Settings → Engram Memory — a standalone first-class settings section
(right after the Plugins section, not a child tab of it). Default
"All workspaces" view shows every workspace's memories/tasks/links
grouped by workspace (dropdown + prev/next workspace pager; the memory
table additionally pages 10 rows per page with a jump dropdown, fixed
column widths, 3-line clamped content ellipsis and full text on hover).
Overview
stat cards (counts by workspace/kind, auto-capture totals, per-workspace
index token estimate, cumulative GC totals), a searchable /
filterable memory table with archive + delete actions, an ESR task board
with an inline "new task" form and a per-task "fill evidence to close"
(artifact / evaluation / memory_ref → STABLE, same gates as esr_close),
a node + relation list (nodes are domain objects the model registers via
esr_node — package/service/repo/concept; relations via esr_link), a separate
tab (hand-rolled force-directed SVG, no chart library so
the bundle stays pure: entities as circles, tasks as check badges, relations
colored per type with direction arrows; drag nodes, pan, wheel-zoom, re-layout,
hover highlights the neighborhood, clicking a node pops a floating panel with
its incident relations and linked objects; dangling links are counted and
warned about), an tab that renders the exact
index block (order 40) and task/closure block (order 41)
the model sees each session — same pure functions as the system-prompt
sections — as two terminal-style panes with per-line coloring (block
headers, task lines, drill hint, and the data-driven reminder
highlighted), line/char/~token cost chips plus memory/task/link/node count
chips, 20s auto-refresh and one-click copy of the raw block text (backed by
the new route). Every ESR task
card (board and ESR tab) carries an — a small
three-arc SVG donut mapping artifact · evaluation · memory_ref, all green
when closure-ready, amber while gapped, gray with no evidence yet; the
board header adds an showing overall evidence
completeness (%) plus how many in-progress tasks are closure-ready. Pure
SVG, no chart library, bundle stays clean. A tab
turns the /stats usage rollup (workspace × day) into a pure-SVG dashboard:
three gauges for (benchmarked against the 0.34 escalate
threshold, amber + hint when low), and , five stat cards (total / esr / memory calls, avg hits per
query, failures), a 14-day mem-vs-esr stacked bar chart and a Top-8 tool
breakdown (mem blue / esr purple), 20s auto-refresh with an automatic
small-sample (<10 calls) warning. A (master–detail)
opens on the right when you click any task / memory / node / relation
row: task cards show state badge, evidence ring, full id + timeline,
gap list, clickable memory references and an inline "fill evidence to
close" form that refreshes the lists on success; memory cards show full
text, tags, signal/hits/TTL and provenance metadata; node cards list all
incident relations (typed, colored, with direction + confidence); clicking
a task's memory reference jumps straight to that memory, with a hint when
it is not in the loaded set. And a
memory-GC panel (dry-run toggle + run button + pointer report). The GUI
create/close use the host's new and
routes. Model-side proactivity is driven
by the [ENGRAM]/[ESR] injected blocks: multi-step work gets a task now,
recurring domain objects get a node, related tasks/nodes get a link.
The browser half is served by DSH's client-module loader directly from this
package (dsh.client + exports["./client"], no web-application rebuild); the
data comes from the loopback-fenced /api/dsh-engram/* route family. The fence
stays closed by default; to reach the memory viewer from an authorized tunnel
hostname, list it in the plugin's trustedHosts config (e.g. via the registry
or a profile patch):
// patch/engram.json
{ "engram": { "trustedHosts": ["cream-club-fragrances-caught.trycloudflare.com"] } }
If you change client/src, rebuild the bundle with:
npm run build:client
Tools
| Tool | Purpose | Kind |
|---|
engram_store | Explicitly store one memory (kind, tags, optional entity anchor, optional supersedes/contradicts ids, optional source/conditions) | write |
engram_recall | Deterministic keyword recall over workspace memories; optional search_sessions FTS over past sessions; opt-in scope=global/recallScope cross-workspace recall with [W:<ws>] origin markers | read |
engram_detail | Full record of one memory id (provenance, tags, hits) | read |
esr_task | Create a task entity (draft → active) | write |
esr_close | Close a task via the evidence protocol (artifact + evaluation + memory_ref) | write |
esr_link | Add a typed relation between two entities (mini graph) | write |
esr_dep | Add a dependency edge (blocks / relates-to / parent-of) between tasks | write |
esr_claim | Atomically claim a task (assignee + claimedAt, draft → active) | write |
esr_unclaim | Release a claimed task's assignee | write |
esr_ready | List claimable tasks (no open blocker, unclaimed) | read |
esr_status | Pull live ESR state + derived hints (since_revision short "unchanged" response) | read |
esr_node | Create/update an entity node (stable symbol) | write |
esr_gc | Run the memory GC for the workspace (dry_run:true previews) | write |
esr_model | Precomputed mental model of the workspace (brief/full, max_chars) | read |
The [ESR] block is a frozen per-session snapshot and never auto-refreshes —
call esr_status for the live truth.
GC: two planes — memory-panel reclamation + Context GC (replacing auto-compact)
Memory-panel reclamation (store maintenance)
A scheduled sweep (gcIntervalHours, default 24h) plus a manual esr_gc /
GUI button keeps the store bounded the pi-esr way — mechanical, working-set
protected, archive-only:
- TTL-expired memories are archived (soft; the id stays, retrievable via the
GUI's archived filter);
- over-cap workspaces evict the lowest-value non-protected memories;
- stable tasks past
gcStableRetentionDays become archived and leave [ESR];
- links whose both endpoints are gone are dropped (dangling edges).
GC never touches the working set: memories referenced by an active task
(memory_refs), task-kind memories, and already-indexed hits
(hits >= promoteHits). Run esr_gc with dry_run: true to preview. Nothing
is hard-deleted — the report ends with re-fetch pointers for everything it
archived, so archives are recoverable, not lost.
Context GC (replaces DSH's auto-compact)
DSH's default context compression is a lossy LLM full summary
(compaction-basic): evicted history is crushed into prose — un-queryable, and
the summary itself burns tokens. dsh-engram's auto-GC replaces it:
ContextGcEngine extends BasicCompactionEngine overrides the single
summarize() hook and swaps the summary body for:
- Scan — the evicted messages are scanned for provenance anchors: memory
ids echoed by
engram_store/engram_recall/engram_detail, tsk_*/ent_*
touched by esr_*, and file_path anchors (esr_gc and other management
calls are NOT anchors);
- Pointer summary — every evicted category carries an explicit re-fetch
call (
engram_detail(id: "…") / engram_recall(query) / the [ESR] block /
esr_ready); the active working set is restated, never evicted;
- Narrative safety net — only un-provenanced turns (pure chat / reasoning)
get a scoped LLM summary (
gcNarrative, default on; off = the whole path is
zero-LLM and un-provenanced turns are kept as truncated verbatim).
Triggers follow DSH's own compaction timing (step pressure / context-overflow /
/compact); the lock, replay validation, tool-call/result pairing and token
pricing are all reused from the basic engine. Any error falls back to the
default compressor — compaction is never broken; with gcReplacesCompaction: false (the default since v0.6, or a host without dsh-compaction-basic) a bare
BasicCompactionEngine is mounted instead, so the compaction service never
disappears while dsh-engram is installed. Mounting registers the compaction
service on the plugin's fiber, so unloading/reloading engram restores the
default engine. Taking over compaction is opt-in: turn on
gcReplacesCompaction (and autoWebCompaction for web presets).
Shrink gate: the harness refuses any checkpoint that is not strictly
smaller than the evicted span (summary is not smaller than the shadowed content) and rolls the compaction back — otherwise the context never shrinks
and the session eventually hits the model-side overflow. Context GC
self-predicts that gate: it mirrors the harness framing with the host
tokenMeter (verified token-for-token identical) and truncates the
pointer/narrative body to the evicted span's token budget (the pointer head
and working set survive; the trimmed tail stays recoverable in the session
log), so compactions actually commit. Full from-scratch usage:
docs/CONTEXT-GC-GUIDE.zh.md.
Where it is enabled (the entry point)
Assembly happens on two planes, but both switches are opt-in (no takeover by
default):
- Host plane (headless / TUI / base profiles):
the main plugin (
lib/index.js) mounts the engine inside ctx.effect via
mountCompactionEngine(ctx, resolved, { readWorkspace }) → new Engine(),
registering the compaction service on its own fiber; the plugin patch
disables the base compaction-basic row, making engram the sole host-plane
provider. gcReplacesCompaction: false (default) → a bare
BasicCompactionEngine (DSH's own summarizer); true → ContextGcEngine
(pointer summaries). Enable via profile patch
(config: { gcReplacesCompaction: true }) or the settings card.
- Preset plane (web profile) — covers every stock preset once enabled: web keeps compaction inside each agent preset's isolated realm
(the shipped presets mount their own
compaction-basic), unreachable from a
host-plane row — and unreachable from a profile patch too (mountPreset
builds its Include.Config with no patches; verified against the harness
source). So on boot, once the agentPresets service is ready, the plugin
auto-edits every preset that still has the stock layout — the default
preset plus the whole roster, across BOTH preset generations: legacy
<root>/<id>/agent.cordis.yml compositions (DSH ≤ 0.1.x, row at 4 spaces)
and the declarative @deepseek-ai/dsh-web-app/presets/<id>.patch.yml files
(DSH ≥ 0.2.0, row at 14 spaces inside the @deepseek-ai/dsh-agent-preset
declaration; a user-defined declarative preset with no shipped file is left
alone). The row is located by YAML structure (a list item carrying
group: true) and validated by both its - id: and name: values, so the
swap is column-position independent: it swaps the compaction-basic row
inside the compaction group for dsh-engram/compaction. That auto-wiring is:
- off by default (opt-in) (
autoWebCompaction: false, toggle in the
settings card's memory-GC section); once enabled it is idempotent — an
already-wired preset is left alone;
what "off" means: the plugin stops auto-wiring on later boots (the
headless/TUI/base host plane is unaffected) — it does NOT un-wire presets
that were already wired; use npm run web-compaction:revert for that;
- per-preset: only stock-layout presets are touched — a preset with a
custom compaction group, or none at all (e.g. the shipped
minimal
preset), is never written, only logged;
⚠️ : once a preset references
, removing dsh-engram leaves every such row dangling
and web sessions hang on load — so run
(restores every preset) BEFORE uninstalling. A harness upgrade that ships a
fresh standard preset overwrites the swap and self-heals the dangling
reference.
First impressions after an npm install (perceive / guide / configure)
- Semi-automatic (opt-in): DSH's own summarizer stays the default. Turn on
gcReplacesCompaction (host plane) and autoWebCompaction (web presets) in
the settings card, then boot once — mountCompactionEngine takes over and
the stock presets are rewritten. No file editing after that.
- Perceivable: at boot the plugin writes an authoritative snapshot to
$DSH_HOME/engram/context-gc.status.json (host plane mode + the per-preset
web results + effective config). At any time run npx dsh-engram status or
npx dsh-engram doctor; the web memory board (ESR kanban) also shows a
header chip ("Context GC·主机 ·N 预设") fed by the overview API. Two startup
log lines confirm it too: engram context-gc: compaction = Context GC … and
engram web-provision: wired Context GC into preset ….
- Configurable: one settings card (memory-GC section) governs everything:
autoWebCompaction (web auto-wiring, default off), gcReplacesCompaction
(Context GC vs host summarizer, default off), gcNarrative (narrative
safety net).
Changes take effect after restarting dsh web, and the web rows are
refreshed to match automatically.
Verify it took: after restarting dsh web (or the profile) —
- host plane (headless/TUI/base): the startup log shows
engram context-gc: compaction = Context GC (mechanical eviction + re-fetch pointers);
- web plane: the log shows one line per wired preset —
engram web-provision: wired Context GC into preset "standard" (…); restart dsh web so sessions pick it up,
or npx dsh-engram status reports wired for each; sessions composed
from ANY preset then compact via mechanical eviction + re-fetch pointers.
A line like … keeping the existing compaction service means another provider
already owns compaction in that realm (the swap didn't take); web-provision … skipped (custom)
means that preset was customized (or has no compaction group) and was left untouched;
dsh-compaction-basic unavailable means the backend is missing and everything
falls back to the default compressor.
BoundaryPrune (subtask-boundary pruning, opt-in since v0.7.11 + ledger veto)
Context GC owns what a compaction summary says; BoundaryPrune owns when a
deterministic prune fires: at todo-completion progress and goal terminal
events, it calls DSH's own toolResultPruner (deterministic surface-replace,
zero LLM) so long-stale over-threshold tool results stop replaying in every
later request. The two switches are independent.
- Evidence chain (oracle analysis +
proposal): 97 real sessions measured —
2.77B–7.53B chars of tool-result replay exposure (~692M–1.88M tokens), 96%
of result bytes from bash/read/run_code; the counterfactual simulation says
todo boundaries cover 79% of that exposure and a repayment gate lifts
cache-aware net savings 12–32% on every threshold config.
- Repayment gate: a prune rewrites the prefix (one KV-cache break), so it
fires only while the session is still expected to run at least
boundaryPruneMinRemaining (default 50) more requests — expectation is a
workspace-history prior with ×2 growth projection past it, so long sessions
are never starved.
- Honest activation surface: sessions with no goal and no todo progress
simply never trigger (no boundary → no prune, zero side effects); the
corpus-wide dry run activates on 38% of sessions. If the pruner service is
absent the mechanism disables itself with one log line.
- Live read (2026-10-26, two days of real use): the Context Ledger recorded
the host pruning 64,200 chars on its own across two sessions (4 prunes in a
220-request session that never wrote a completed todo), while engram's own
leg produced 7,841 chars from a single boundary prune. The chain itself works
(v0.7.7–v0.7.9 fixed gate/execution/accounting); the coverage does not.
- Boundary sources: todo-progress completion and goal terminal events
(default), plus an optional request-count cadence
(
boundaryPruneIdleCadence, default 0/off) — the "step idle" source that
covers the ~62% of sessions which never touch todo/goal (recon:
eval/recon-idle-boundary.mjs; +85% saved chars at cadence 50 in the
counterfactual eval/simulate-idle-cadence.mjs), still repayment-gated.
- Threshold override:
boundaryPruneThresholdChars (default 0 → host's
8192). Set e.g. 2048 with smaller retained edges
(boundaryPruneHeadChars/boundaryPruneTailChars) to prune earlier —
T2048+small-edges lands 2.4× the first-prune savings of the host default on
the real corpus (10.14M vs 4.23M chars). Lowering the threshold is safe with
resultPack on, which archives the removed middles.
- Default off since v0.7.11:
boundaryPrune: true is the opt-in. Fifteen
hours of live ledger proved the host prunes over-threshold results on its own
(64,200 chars across two sessions, zero engram involvement) and that the
boundary source never fires in an ordinary 220-request session — only planning
sessions that use todos/goals get anything out of it. Adaptive gate
lets the Context Ledger stop further fires once
this workspace's own accounting is net negative (≥3 fires) — evidence can only
turn pruning off. The log shows
.
ResultPack L2 (archive + result_recall, default ON)
BoundaryPrune makes pruning lossless: the removed middle of every pruned
tool result is redacted (same rules as the memory write path) and archived to
~/.dsh/engram/resultpack/<callId>.json, and the pruned surface node gets a
tail handle — [middle archived: result_recall(<callId>) 可取] — so the model
knows the full result is one step away. result_recall(callId, offset?)
returns the archived middle in paged text blocks (progressive disclosure, the
same philosophy as [ENGRAM] → engram_detail, and the complete
ObservationPack pattern from SoL-Pi mechanism ③).
- Lossless by construction: L1 pruning shrinks the replay; L2 keeps the
original recoverable on demand — the "model rereads a pruned result" failure
mode goes to zero.
- Scoped cost: archives live outside the memory domain (big blobs never
bloat the domain JSON or the [ENGRAM] index); retention is a mtime sweep
(
resultPackRetentionDays, default 7) riding the same GC cadence.
- Self-disabling: L2 without
boundaryPrune has nothing to archive, so it
registers no tool and emits one warning.
- Default off:
resultPack: false (must be on together with
boundaryPrune: true). When active the log shows
engram result-pack enabled: result_recall(callId, offset) registered.
Activation telemetry (SoL-Pi M15/C13)
Mechanisms must be provably activating, and dormant ones are off by default.
/api/dsh-engram/stats (and the GUI overview) now answer
activation: { boundaryPrune, recall }:
boundaryPrune — cumulative, durable record of real BoundaryPrune fires
(fires / prunedResults / charsRemoved / byKind todo|goal|step / lastFire),
written to ~/.dsh/engram/boundary-prune.status.json on every landed prune,
independent of resultPack.recall — the injection leg of the recall funnel: how many [RECALL]
blocks actually made it into sessions (recall-injections.json), closing
the loop already measured from session logs (engram_recall queries → hits →
engram_detail follows) so you can see injection→query→detail activation.
Context Ledger (net-savings accounting)
Turns "saves tokens" into a checkable number. Every model request is one ledger
tick: injected [ENGRAM]/[RECALL] and [ESR] blocks are charged per request
they survive; results removed by BoundaryPrune are credited per later request
they no longer replay; every prune pays one prefix-rewrite penalty
(1 − ledgerCacheDiscount) (the KV-cache break). GET /api/dsh-engram/stats
and /overview return ledger:
totals — requests / injectedChars / avoidedChars / rewriteCount / rewriteChars / netChars / netTokens with
net = avoided − injected − rewriteChars × (1 − discount);byWorkspace / byDay — the same arithmetic per workspace and calendar day
(byDay newest first);openSessions — sessions still accumulating in memory (a restart loses at
most flushEvery requests).
The Telemetry tab renders it directly (net-token headline, saved / injected /
rewrite-penalty cards, a 14-day red/green net chart), and the memory overview
adds a "net context tokens" stat card. Durable rollup at
~/.dsh/engram/ledger.json (atomic write, daily buckets, delta flush); the hot
path is an integer add, zero I/O and zero model calls.
Defaults: ledgerEnabled: true, ledgerCacheDiscount: 0.1; off = no listeners,
no writes.
Two credit legs (v0.7.1 — before this the ledger only charged costs, so the
net was structurally negative):
- L1 boundary prune — chars removed by
toolResultPruner, credited per
later request. Requires a host-plane toolResultPruner; DSH ≥ 0.2.0 web
profiles move that row into each agent preset's isolate realm, so the bundle
patch re-enables the host row — check boundaryPrune.armed in the status
snapshot.
- Compaction (Context GC) — every checkpoint credits
span chars − checkpoint chars from the engine that produced it; host plane and the web
preset realm (bridged through the engramLedger service) write the same
ledger.
Honest scope: tokens are chars / 4 estimates, not provider billing, and no
quality delta is claimed. Until a workspace actually prunes or compacts its
"saved" leg is 0, so the net equals injected + rewrite penalty (negative is
expected) — the Telemetry tab says so inline.
JIT push (v0.6-C)
Pull-only recall measured dormant (24 engram_recall / 8 engram_detail across
112 real sessions), so v0.6 pushes on three deterministic events instead. One
line per event, at most 3 lines per session (jitMaxLines, jitMaxLineChars=220),
and silence when there is nothing relevant to point at:
| Event | Trigger | What is pushed |
|---|
| Error recurrence | the same cmd: signature fails a second time and a prior error/procedure memory exists | that memory's id + a one-line excerpt + engram_detail |
| File focus | the same file is edited twice in a row and a filePath/dir-anchored memory exists | preference first, then fact/decision |
| Goal completion | goal complete/block while active ESR tasks remain | the first two open tasks + esr_status |
- Append-only: a pushed line is never edited or withdrawn, so the prefix
changes once per push. Each push charges a rewrite
(
estimatePrefixChars + line length) in the Context Ledger
(rewriteCount / rewriteChars), visible on the telemetry page.
- Silent without a hit — no memory/task to point at, no line.
- Zero LLM, pure rules;
jitEnabled:false disables it, and the empty
[JIT] section costs nothing.
- Not yet: the resume brief (next-day continuation) and an explicit
remaining-requests ROI gate — today the 3-line budget plus the ledger are the
guardrails.
Session episode + resume brief (v0.7-A/D)
- episode — at
session/disposed, a zero-LLM digest of the event log: files
touched, commands run, failures that later recovered (cmd×N→fixed), written
commit subjects, and still-open todos. Written only when there is substance (a
bare ls session produces none); never indexed; the newest 20 per
workspace are kept and older ones archived.
- resume brief — at the first assembly of a new session, the newest episode
from the last 7 days is injected as a
[RESUME] line (≤240 chars) plus an
engram_detail <id> pointer, frozen with [ENGRAM]/[RECALL]. No recent
episode → no line; returning to a project does not need a re-briefing.
- Settings:
episodeEnabled, resumeBriefEnabled (both default on), plus
episodeMaxPerWorkspace / resumeBriefMaxAgeDays / resumeBriefMaxChars.
Host requirement (silent-dormancy warning). Episode, resume brief, [RECALL]
and the session-end todo sink all read the session's event log. On ≥0.2.x hosts
the log arrives through ctx.sessionQuery.live(), session.snapshotEvents() or,
as a last resort, session.seq + session.eventAt(seq). A host that exposes
none of these keeps the three features dormant: memory storage, engram_recall
and engram_detail still work, but nothing is injected at session start and no
episode is written. The plugin logs one engram: this host exposes no readable session event log … Session surface: … warning per session naming what the
session object does offer — that line is the thing to paste into a bug report.
Known case: the desktop runtime (@deepseek-ai/dsh-desktop-runtime) on
Windows, tracked as issue #9.
dsh-engram status also prints a push-legs line — 推送腿：事件日志读法 sessionQuery · 最近 … — the read path the live process actually took, written
to $DSH_HOME/engram/push-legs.jsonl. A dormant host shows
事件日志不可读（none）… 休眠中 there, so the four features can be checked
without hunting a startup log.
Entity pages + entity cards (v0.7-B)
esr_node grows from a dot on the graph into a readable page: current
truth, timeline, open loops, conflicts and relations for one thing
(package / repo / service / bug…).
- Entity card (in the index, default on) —
[ENGRAM] renders ONE line for
each active entity (anchored memories or open tasks), at most 2:
[ENT] dsh-engram · package · 3 mem · 1 open · truth: … #id · page: engram_detail(entity="dsh-engram")
Those entities' flat memory lines are not repeated; other entities stay in a
single nodes: line.
- Full page (on drill) —
engram_detail(entity="<name|id>") returns
truth / timeline / open loops / conflicts / relations, with memory ids and
evidence gaps. A name with anchored memories but no registered node still
gets a virtual page.
- Pure, deterministic: preference leads
truth, superseded rows never enter
it, conflicts are de-duplicated by unordered pair (#a vs #b).
- Settings
entityCardsEnabled (default on); entityCardsMax /
entityCardChars tunable.
Revalidatable anchors + layered TTL (v0.7-C)
"Treating the past as truth" is the strongest credibility critique against
memory plugins. v0.7 gives file-backed memories an anchor:
- on write, record
{path, mtimeMs, size, hash?} (sha256 for files ≤256KB,
stat only for larger ones) in the memory row; relative paths resolve against
the workspace;
- at session start a bounded revalidation (≤40 rows) tags changed/missing
sources
stale; the [ENGRAM] index and entity truth stop showing them,
but nothing is deleted — engram_detail still returns them and the entity
timeline marks · stale;
- layered TTL:
preference / procedure permanent; episode 14 days;
error 60 days; everything else expireDays (180). Hard deletes still do not
exist — GC only archives.
- Settings:
anchorRevalidate (default on), episodeTtlDays, errorTtlDays.
- Not done: the offline consolidation job (episode abstraction/merge) —
deterministic rules can only roll up, not abstract semantically; deferred
until v0.8 alongside playbooks, rather than shipping a lossy pseudo-merge.
Auto-capture policy
Capture is deterministic and offline — it only sees tool results, never the
conversation. Exactly what earns a memory record:
| Tool result | Action | Signal |
|---|
git commit … -m "subject" | record — the written subject is the memory | 0.55 |
git merge / rebase / cherry-pick / tag / checkout -b | record (milestone) | 0.5 |
git push / git stash / commit without -m | skip — plumbing echo, not a decision | — |
| write/edit of a significant config & doc path | record | 0.3 |
| read of a config path | record | 0.3 |
| repeated tool error | record (deduped by message) | 0.25 |
Explicit engram_store writes are always recorded regardless of these rules
(rate-limited per session).
Who earns a [ENGRAM] index line (this is what actually touches the prompt):
signal >= minIndexSignal or hits >= promoteHits or kind === "task",
then capped by indexMaxLines / indexMaxChars.
Procedures are evidence-gated (v0.8-A). An auto-captured fixed: <cmd> row is
a hypothesis, not a lesson: it starts with evidenceCount = 1 and stays out of
the index until the same command is verified 3 times with the last
verification younger than 14 days (procedureMinEvidence /
procedureMaxAgeDays). The moment that command fails again the row is
invalidated (invalidatedBy = the failing memory's id, signal → 0.2) and leaves
the index immediately — a just-broken procedure is worse than none, because the
next agent would copy it verbatim. Hand-written procedures with no evidence
fields keep the older hits >= promoteHits path.
Promoted procedures can become real Skills (v0.8-B). An index line is a
hint; a DSH Skill is an artifact the host discovers, advertises in the session
catalog, and any agent (or subagent) can load in full:
npx dsh-engram skill export # → <workspace>/.dsh/skills/<name>/SKILL.md
npx dsh-engram skill export --dry-run # show the plan, touch nothing
Only procedures that pass the evidence gate are exported (≥3 verified
successes, verified within 14 days, not invalidated). Each file carries a
generated-by marker: the exporter rewrites and eventually deletes only its
own files, so a hand-written skill sharing the root is never touched. When a
procedure stops holding, its skill file disappears on the next export — the host
catalog must not keep advertising a fix that just broke. npm run skill:check
proves the seam end-to-end by loading DSH's own dsh-skill-filesystem provider
and listing the exported skill through it.
The approval queue is in the GUI too (v0.8.4). The memory tab shows
"Playbook 待批准" with evidence counts and approve/revoke buttons, backed by
GET|POST /api/dsh-engram/approvals. The route recomputes the list with the same
planner skill export uses, so a click can never approve something the CLI would
skip; revoking returns a row to pending rather than deleting its evidence.
Writing a skill needs a human yes (v0.8-D). Evidence (≥3 verified
successes) gets a procedure into the private index line; it does not, on its own,
put a file where the host advertises it to every agent. skill export therefore
writes only approved procedures — skill pending lists what is waiting,
skill approve <id> / skill revoke <id> maintain the record in
~/.dsh/engram/playbook-approvals.json (kept out of the memory store: an approval
is a human act, and a pool must not approve itself), and --force is the audited
escape hatch. Three successful runs prove a command worked; they do not prove the
instruction is sound.
Subagent reports become durable capsules (v0.8-C, advisory). A child's final
report above capsuleMinSourceChars (1500) is distilled by structure — headings,
bullets, question marks, next-step cues — into a handoff memory (结论 / 证据 /
未决 / 下一步) whose footer records source → capsule chars. It does not shrink
what the parent reads: DSH emits tools/result to observers only ("without
exposing a mutation or error channel into the outcome"), so substitution needs a
host seam that does not exist yet. What you get today is durability — the
conclusion survives the child session and compaction — plus a real measurement of
the saving for the day that seam lands. One extra guard keeps the pipe
clean: auto-captured git command echoes — text that embeds a shell chain
(git push: cd … && …) — stay out of the index even when above the signal
threshold, until recall hits have promoted them. Everything else sits quietly in
storage, reachable on demand via engram_recall / engram_detail — "retrieved ≠
injected".
Injected blocks
What the model actually sees (rendered once per session, then frozen):
ESR operating protocol (static, byte-identical every turn)
1. Before starting: esr_ready to see claimable work; esr_status for live state.
2. Multi-step work → esr_task (draft first); claim → esr_claim; finish → esr_close (all three gates).
state is the single source of truth: unsure → call esr_status.
[ENGRAM] workspace: symbolic-index · 2 memories · 1 task(s) active · 0 links
[D] 06-18 Decided: use sqlite-vec for retrieval #a2331d87
[T] 06-18 Retrieval upgrade — ACTIVE · gap: artifact, evaluation, memory_ref #tsk_8b26
drill: use [RECALL] below · engram_detail <id> (full record) · engram_recall <query> (more) · esr_task/esr_node/esr_link (work)
[RECALL] recall · 1 hit(s) · first msg: retrieval
- [D] 06-18 Decided: use sqlite-vec for retrieval upgrade #a2331d87 ×2
context: use above · engram_detail <id> · engram_recall <query>
[ESR] tasks: 1 active / 1 stable
- tsk_0d: Retrieval upgrade — ACTIVE · gap: artifact, evaluation, memory_ref
- closed: tsk_9a (RAG eval) · +1
snapshot from session start — WILL NOT auto-refresh; call esr_status for live state
# this-session actionables (frozen)
promote: 2 pending todo(s) vs 1 ESR task(s) — esr_task(name="…") #suggest-promote
Prefixes: [D] decision · [E] error · [P] procedure · [F] fact ·
[I] insight · [H] handoff · [T] task. A procedure re-used enough times
(hits >= promoteHits) becomes a "proven experience" — its prefix upgrades to
[P✓] and it is stably ranked first in the block (newest-first within each
group, keeping the block deterministic): our zero-LLM counterpart to TencentDB
Agent Memory's "Skill = validated, executable experience". Membership follows
the Auto-capture policy (signal threshold / recall promotion / git-echo guard),
bounded by the configured line and character budgets. # ids address the full
records via engram_detail. When a workspace has no tasks, [ESR] still renders
one line naming esr_task/esr_close so the mechanism stays visible to the
model instead of vanishing.
[RECALL] (session-start auto recall, autoRecallOnStart) runs one deterministic
BM25 pass over the workspace keyed on the session's first user message and injects
the top autoRecallLimit hits, bounded by the autoRecallMaxChars character
budget; nothing renders on zero hits or an empty workspace. It is a pointer — the
model should treat it as "this is relevant; engram_detail for the full record" —
rather than full-text dumping. It is a pure read (never bumps hits), superseded
stale truth never occupies a slot, and it freezes per session with [ENGRAM].
Config
Defaults are token-conscious; override any key via the profile patch
(~/.dsh/profiles/web/cordis.patch.yml) or the web config card:
- id: engram
config:
autoCapture: true # zero-LLM tool-result capture
autoCapturePreferences: true # zero-LLM user-correction/preference capture (prefer/avoid/switch-to cues)
preferenceCapturePerSession: 8 # per-session preference memory cap
boundaryPrune: true # ResultPack L1: prune stale over-threshold tool results at todo/goal boundaries (default ON)
boundaryPruneAdaptive: true # ledger veto: stop pruning this workspace once its net is negative (>=3 fires)
boundaryPruneThresholdChars: 0 # threshold override (0 = host 8192; e.g. 2048 prunes earlier)
boundaryPruneHeadChars: 0 # retained head under an override (0 = host 4096)
boundaryPruneTailChars: 0 # retained tail under an override (0 = host 1024)
boundaryPruneIdleCadence: 0 # request-count boundary for todo/goal-less sessions (0 = event-only)
resultPack: true # ResultPack L2: archive pruned middles + result_recall paged recall (default ON; needs L1 on)
resultPackRetentionDays: 7 # archive retention days (mtime sweep on the GC cadence)
sessionSearch: true # engram_recall may also FTS past sessions
recallScope: workspace # recall scope: "workspace" (strict isolation, default) | "global" (opt-in cross-workspace recall with [W:<ws>] origin markers)
autoRecallOnStart: true # session-start auto recall (false = back to pure pull)
autoRecallLimit: 3 # [RECALL] max injected hits (few and precise)
autoRecallMaxChars: 700 # [RECALL] char budget
autoCapturePerSession: 40
indexMaxLines: 12 # [ENGRAM] line cap
indexMaxChars: 700 # [ENGRAM] char cap (token budget)
minIndexSignal: 0.4 # auto-captures below this stay out of the index
# (git command echoes are excluded regardless,
# until promoted by recall hits)
promoteHits: 3 # ...until recalled this many times
procedureMinEvidence: 3 # verified successes before a `fixed:` procedure
# earns an index line (one success proves nothing)
procedureMaxAgeDays: 14 # ...and the last verification must be this fresh
expireDays: 180 # memory TTL (0 = never)
maxMemoriesPerWorkspace: 2000
gcEnabled: true # scheduled memory GC
gcIntervalHours: 24 # sweep cadence
gcStableRetentionDays: 120 # stable tasks leave [ESR] after this
gcReplacesCompaction: false # Context GC: take over DSH auto-compact (default off = host summarizer; true = opt-in pointers)
gcNarrative: true # scoped LLM narrative for un-provenanced turns; false = fully mechanical (zero LLM)
autoWebCompaction: false # rewrite stock web presets (default off; explicit opt-in, revert with the CLI)
jitEnabled: true # JIT push: error recurrence / file focus / goal completion → one memory line
episodeEnabled: true # session episode: zero-LLM digest at teardown (never indexed)
resumeBriefEnabled: true # resume brief: newest episode from the last 7 days at session start
entityCardsEnabled: true # entity card: one truth/tasks/drill line per active entity in the index
anchorRevalidate: true # re-stat filePath anchors at session start (changed/missing → stale)
episodeTtlDays: 14 # episode TTL (preference/procedure permanent; error 60d; others expireDays)
ledgerEnabled: true # Context Ledger: injected cost / prune savings / rewrite penalty → net tokens
ledgerCacheDiscount: 0.1 # cached-prefix price share; rewrites pay (1−discount)
engramIndexOrder: 40 # systemPrompt section order (before tools band)
esrOrder: 41
Development
npm test # 381 tests: core + web API + GC + usage-observability + Context GC + ESR triggers + BoundaryPrune + preference learning + Context Ledger + JIT push (node:test)
npm run eval # offline recall + structure benchmark (deterministic)
npm run build:client
Testing & evaluation (two layers of real testing)
npm run eval (eval/recall-bench.mjs) is the deterministic layer —
LongMemEval-style: a controlled corpus (ASCII + CJK, known tags/entities/
timestamps) measured through the real store/recall path (openEngramDomain +
domain.recall), reporting Precision@k / Recall@k / MRR / Hit@1 per probe
(exact tag, substring, multi-term, CJK, phrase-single, ordering, negative)
plus StructMemEval-flavoured structure metrics (exact-duplicate dedup rate,
entity anchoring coverage, dangling-link hygiene). Honest, reproducible
numbers — everyone gets the same output. Current run: AVG P@k 0.770 /
AVG R@k 1.000 / MRR 0.889 / hit@1 0.889 / 0 negative false-positives /
dedup 1.0.
/api/dsh-engram/stats + the observability panel is the real-session layer —
it answers how the model actually uses the memory in production (ESR
proactivity ratio, recall hit rate, detail conversion), while the eval
answers how good the retrieval layer itself is. The stats read the
concatenated multi-frame session.jsonl.zstd logs with Node's built-in
node:zlib (frame scan + one reused decoder handle) — no system zstd CLI
is required, so the telemetry works out-of-the-box on Windows and other
hosts without an external codec.
Repo layout: lib/ (host half: store / capture / index-block / tools / api /
settings), client/ (browser half, TSX + build.mjs), test/ (node:test).
Troubleshooting
The web GUI opens and immediately shows “Failed to load plugins”, with a
loader error like:
failed to apply loader entry … (@linxin666/dsh-client-ui-web-ui-settings):
keyed slot "settings.plugin.item" requires options.key
Cause: DSH hosts since 0.1.0-rc.7 declare the config-card slot
settings.plugin.item as keyed by the settings namespace a card edits —
which is exactly how dsh-engram's own config card registers (under
key: "dsh-engram"). The @linxin666/dsh-web-ui-all family before 0.2.0
(dsh-client-ui-web-ui-settings) registered its group card into that slot
without a key, and because a single failed loader entry aborts the whole
boot, the GUI stays stuck on the failure page.
Fixes:
- Proper fix — upgrade the family:
@linxin666/dsh-web-ui-all@^0.2.x.
The 0.2 line moved its settings surface out of the keyed slot into a
first-level settings.section (the upstream fix for exactly this error).
- Immediate unblock: add
key: "web-ui-plugins" to that one
settings.plugin.item registration in the installed
node_modules/@linxin666/dsh-client-ui-web-ui-settings/lib/client.js, then
restart dsh web. (Under namespace-keyed dispatch the group card simply
stays hidden; nothing else on the page is affected.)
Related
- symbolic-index — the original cross-session
memory plugin (5-signal RRF fusion, sqlite-vec, Dream Engine).
- pi-esr — project-lifetime evidence-driven
task states; the closure protocol here is its lite form.
License
MIT
