Graph Memory
Graph Memory for DeepSeek Harness, compatible with OpenClaw
Bound the context. Keep the memory.
A native DeepSeek Harness memory plugin that keeps recent conversation turns, archives older history, and recalls exact source-backed knowledge when it matters.
中文 ·
dsh.so ·
20-turn benchmark ·
Upgrade guide
The problem it solves
Long agent history becomes graph navigation plus a compact recent-turn context
Graph Memory owns the model-visible historical surface without deleting DSH's event log. By default it keeps the newest five completed user turns, removes completed reasoning/tool traces from future requests, and recalls relevant older or cross-session source Q/A automatically.
The 1.6 turn-memory navigation upgrade
| Before | Now |
|---|
| Extract TASK / SKILL / EVENT directly from messages | Create one self-contained turn summary, then derive SPO from that same sentence |
| Graph nodes could become the factual payload | Summary, SPO, and communities only navigate; original question and final answer remain the evidence |
| Old memories from the active session could be filtered wholesale | Exclude only sources still visible in the fresh window; archived same-session and cross-session recall share one path |
| Community expansion could pull a whole neighborhood | Local LPA narrows candidates, query-time PPR ranks them, and only matched Q/A is recovered |
| DSH retained complete tool and reasoning traces | Completed turns retain question + final answer; older prefixes collapse to one fixed marker |
Writing one completed turn costs exactly one auxiliary LLM call. Community detection and PPR are local. There are no hard-coded node/edge counts, semantic direction gates, or JSON repair that turns invalid output into accepted data. Read the complete design, source map, and porting sequence →
Upgrading to the on-by-default decay curve
Decay maintenance is on by default: the engine auto-deprecates silent peripheral nodes after autoDeprecateAfterDays (30d, disconnect-only) and hard-deletes decay-deprecated nodes after purgeAfterDays (60d). Unlike messageRetention (which never deletes), this is the one surface that removes data. If you are upgrading from a build before decay and want no auto-disconnect or hard-delete, set any of:
{ "decay": { "purgeAfterDays": 0 } } // never hard-delete (escape hatch; deprecate still runs)
{ "decay": { "autoDeprecate": false } } // no silent auto-disconnect
{ "decay": { "enabled": false } } // disable decay entirely (hard-delete still runs independently)
Only nodes deprecated by deprecated_by='decay' are ever hard-deleted; manually deprecated or freshly merged nodes are untouched, and the revive path re-activates a decayed node once it is referenced again. A non-zero tick logs decay activity to info. Full decay design and knobs →
Measured first
DSH 20-turn first-request context comparison
| Real 20-turn GLM-5.2 run | Historical native DSH baseline | Latest Graph Memory | Change |
|---|
| T20 first request | 56,998 tokens | 11,008 tokens | −80.69% |
| T20 model-visible messages | 171 | 21 | −87.72% |
| T01–T20 first-request context | 532,451 tokens | 165,896 tokens | −68.84% |
| All measured tokens¹ | 2,487,776 | 2,327,728 | −6.43% |
¹ The latest candidate includes 166 main-agent requests, 20 turn extractions, and 41 embedding requests; the historical baseline made 77 main requests. DSH commits and nondeterministic tool loops differ, so this is not a simultaneous strict A/B. First-request context is the direct takeover metric; the full bill remains visible.
20/20 scenario turns passed · 20/20 structured extractions succeeded · 0 quarantined · 20 turn summaries · 92 SPO triples · 30 communities · 20 summary vectors. T11, T19, and T20 automatically recalled out-of-window memory with exact source question and final answer.
Read the Markdown benchmark, per-turn data, method, and limits →
Memory survives the context window
Graph Memory active in DSH
Cross-session recall in a fresh DSH session
The graph is a navigation layer, not a replacement for evidence. TASK, SKILL, and EVENT nodes point back to the original user question and final visible answer; recalled context includes those exact source messages.
Install on DeepSeek Harness
Node.js 22.13+ · no DSH fork · the npm registry already serves @multilab/graph-memory (npm view currently reports 2.0.4), so install from npm:
npx @deepseek-ai/dsh plugin --profile web add @multilab/graph-memory
npx @deepseek-ai/dsh --profile web --dump-config
npx @deepseek-ai/dsh web
To pin a GitHub release instead, use the upstream tag v1.6.0-beta.17 — a real tag on github.com/adoresever/graph-memory (verified via git ls-remote; upstream tags end at v2.0.0, the 2.0.x npm line is not re-tagged there):
npx @deepseek-ai/dsh plugin --profile web add github:adoresever/graph-memory#v1.6.0-beta.17
1.6.0-beta.17 adds DSH 0.1.7-rc.2 (session format V4) compatibility: every durable Graph Memory surface write now uses the producer-owned source kind plugin:graph-memory that V4 admission requires (#113). Older DSH releases accept the same shape.
The npm registry also still serves the old unscoped graph-memory@1.5.8; do not use it to validate DSH — the scoped @multilab/graph-memory package above is the maintained release line.
Confirm that @multilab/graph-memory/dsh is active under Settings → Plugins. The default database is $DSH_HOME/graph-memory/graph-memory.db, normally ~/.dsh/graph-memory/graph-memory.db.
Entry points. The package ships two host-specific defaults: exports["."] / ./dsh resolve to dist/dsh.js (the native DSH/Cordis adapter), while the legacy main points to dist/index.js (the OpenClaw plugin). These are intentionally different hosts: DSH consumers use the @multilab/graph-memory export, OpenClaw loads index via its declarative manifest. Do not change main to dsh.js — it would break legacy resolvers that import the package by name.
Version numbers. package.json (2.0.4) is the npm release line, and openclaw.plugin.json mirrors it exactly — one package, one version. The 1.6.0-beta.x strings in this file refer to the upstream GitHub tag line (v1.6.0-beta.15 through beta.17 exist as real upstream tags, verified via git ls-remote); they are inherited release docs, not this package's version.
DSH Settings Panel
Graph Memory ships a settings UI for DSH: open Settings → Plugins → graph-memory in the DSH web panel. The panel is driven by a schemastery Config schema, so invalid values are rejected inline instead of at runtime.
Hot-effective vs restart-effective
Most fields apply immediately — the plugin listens for DSH's loader/volatile-update event and re-reads the current effective config without a restart:
| Applies immediately | Restart required |
|---|
extractionEnabled, recallEnabled, recallMaxNodes, semanticScoreThreshold | dbPath (reopens the database) |
maintenanceInterval, freshTurnCount, assistantTools, messageRetention, decay | embedding.baseURL / model / dimensions (rebuilds the embedding function and vector fingerprints) |
extract.*, llmReasoningEffort, contextCompactionEnabled, projectCompletedTurnTools | llmMaxTokens (startup snapshot) |
extractionSource, llmProvider, llmModel, extractionModels | |
Extraction model sources
| Source | Meaning |
|---|
session (default) | Follow the current session's provider/model route (zero-config fallback) |
preset | Explicit llmProvider / llmModel pair (backward compatible with the old env config) |
pool | extractionModels list tried in order; a failed model falls through to the next; empty pool cascades to preset, then session |
Embedding API key (credential-store-first)
The embedding API key is never stored in plugin config. The only channel is the apiKeyEnv credential reference (default GRAPH_MEMORY_EMBEDDING_API_KEY, provider-neutral): put the key under that reference name in the DSH credential store (.credentials.yaml refs) or the environment, and the plugin resolves it at connect time through ctx.credentials.resolve. Switching embedding providers changes only the stored value — the reference name stays.
The schema-injected default reference alone does not count as "configured" (no probe noise for unconfigured users); set baseURL (or a custom reference name) to enable vector recall — setting only model does not enable it.
The config panel's Embedding section includes a Test button: it fires one real /embeddings call with the current form values (including unsaved drafts), reporting dimensions, latency, and credential source. The key is resolved host-side via ctx.credentials.resolve — never echoed, never sent to the browser. The per-request timeout is the panel-adjustable embedding.testTimeoutMs (1000–300000, default 30000); slow cross-border endpoints that previously tripped a hardcoded timeout can raise it in the same form, and the credential save's post-save key probes share the same value. Production embedding calls have their own panel-adjustable baseline, embedding.requestTimeoutMs (1000–300000, default 10000, restart to apply): single-text calls use it directly, and batch calls use max(requestTimeoutMs, 2s×texts) so an explicitly raised baseline only lifts the adaptive floor.
Multi-key failover & panel-managed credentials (2.0.14)
Multi-key format. Store several keys under one reference as a comma-separated string — "k1,k2,k3" — in the credential store (.credentials.yaml refs) or the environment variable. The host splits on comma, trims each segment, and drops empty segments, so " k1 , , k2 " resolves to [k1, k2]. A single key (or none) needs no comma and behaves exactly as before. After a save, all keys are probed in parallel so each key's validity, dimensions, and latency show up individually in the panel's report rows.
Failover semantics. When two or more keys are configured, any failure rotates to the next key immediately — no retry within a single key when ≥2 keys are present (a 429/5xx does not back off; it rotates at once). A sticky pointer remembers the last key that succeeded and starts the next call there, wrapping around the ring. When only one key (or none) is configured, the original retry-with-backoff path applies unchanged (single-key zero regression).
Panel management. Settings → Plugins → graph-memory → credential section offers a password input plus Save and Clear buttons. Save writes the whole comma-separated string to .credentials.yaml through the host credential service under the configured reference name; Clear removes the ref. The input value never enters plugin config, never echoes to the client, never enters logs — it travels only to the host credential service.
Effective boundary. Key-value changes take effect without a restart (the credential is resolved per batch and the host watches the credential store), so rotating to a new key or adding/removing keys is hot-effective. baseURL, model, and dimensions remain restart-effective — they are connection-period parameters that rebuild the embedding function and vector fingerprints.
Env-masked references. If the reference name is supplied by a startup environment variable, the credential service reports writable: false; the panel then shows a read-only hint and disables Save, because the key is owned by the shell/environment rather than the credential store. Manage such keys in your shell/environment instead.
All-fail error. When every configured key fails, the embedding engine reports [graph-memory] Embedding API: all N API keys failed (last: <rotation-order last failing key's error>); the per-key breakdown (status, dimensions, latency, error) is shown in the panel's report rows. On startup, an all-fail probe degrades to FTS5 full-text recall without crashing.
Migration
Existing GRAPH_MEMORY_* environment-variable users need zero changes — the env fallback still applies inside apply(), and the default apiKeyEnv reference already points at GRAPH_MEMORY_EMBEDDING_API_KEY. If you previously filled the panel's embedding.apiKey field (≤ 2.0.7, stored plaintext in the profile patch), move the key into the credential store under the default reference name, restart, and delete the leftover embedding.apiKey line from your profile patch — the stale value stays in that file until you remove it. From 2.0.13 onward the plugin rebuilds the embedding section from a field whitelist at the apply() entry, so the legacy apiKey (and any other unknown leftover) is never read into memory; the schema itself does not strip unknown keys. The shipped cordis.patch.yml is slimmed to dbPath only.
Behavior change for existing users: with the default extractionSource: "session", extraction now follows the current session's provider/model route, and a configured llmProvider/llmModel pair is only a fallback when no session route exists. Previously the dedicated pair took priority. If you rely on extraction always using your dedicated model, set extractionSource to preset in the panel (or GRAPH_MEMORY_LLM_PROVIDER/GRAPH_MEMORY_LLM_MODEL behavior is preserved only via the preset source). gm_status now reports the active source.
Troubleshooting: corrupt database quarantine
On startup (and only then), Graph Memory opens its SQLite store; if the file is page-level corrupt (SQLITE_CORRUPT, classic message database disk image is malformed — typically after an abnormal shutdown or disk error), the plugin does not crash. Instead it:
- Renames the corrupt file — plus its
-wal/-shm sidecars — to <dbPath>.corrupt-<timestamp> next to the database (never deletes; the original is preserved for recovery tools).
- Creates a fresh empty database at the original path and keeps running, so the plugin stays active (extraction/recall continue) but memories start over from an empty store.
How you notice: the host log gets one console.warn line at quarantine time, and gm_status reports DB quarantine: N isolated corrupt file(s) … whenever quarantined leftovers exist next to the store.
Recovery: the quarantined files stay untouched until you delete them. Try sqlite3 <dbPath>.corrupt-<stamp> ".recover" | sqlite3 recovered.db, then point dbPath at recovered.db (or move it back) and restart. Deliberately narrow detection: other damage shapes (for example SQLITE_NOTADB, "file is not a database") are not auto-quarantined and still raise normally.
What ships
| Capability | Implementation |
|---|
| Context takeover | Configurable newest-N completed turns; one archive marker replaces the older model surface |
| Lightweight extraction | Only the user question and final answer; strict structured tool contract; no reasoning/tool transcript ingestion |
| Query-first recall | Vector Top-K with FTS5 fallback; exact source Q/A travels with graph hits |
| Durable memory | Local SQLite, stable provenance, cross-session and cross-project recall |
| Failure behavior | Invalid extraction is quarantined; foreground conversation continues; bad data is not repaired or persisted |
| Host support | Native DSH/Cordis adapter; maintained OpenClaw Context Engine adapter |
Optional embeddings
Graph Memory supports OpenAI-compatible embedding endpoints. Without embeddings it falls back to FTS5 and does not block conversation.
export GRAPH_MEMORY_EMBEDDING_API_KEY='replace-with-your-key'
export GRAPH_MEMORY_EMBEDDING_BASE_URL='https://dashscope.aliyuncs.com/compatible-mode/v1'
export GRAPH_MEMORY_EMBEDDING_MODEL='text-embedding-v4'
export GRAPH_MEMORY_EMBEDDING_DIMENSIONS='1024'
dsh web
DSH tools and extraction route
| Tool | Purpose |
|---|
gm_status | Store, extraction, recall, vector, and retention state |
gm_search | Explicit graph-memory search |
gm_record | Deterministically persist a TASK, SKILL, or EVENT |
gm_stats | Graph and retention receipts |
gm_maintain | One bounded maintenance tick |
gm_retry_extraction | Explicitly retry quarantined extraction |
Automatic recall needs no tool call. Extraction may use a dedicated model via GRAPH_MEMORY_LLM_PROVIDER and GRAPH_MEMORY_LLM_MODEL; optional reasoning and output controls are GRAPH_MEMORY_LLM_REASONING_EFFORT and GRAPH_MEMORY_LLM_MAX_TOKENS.
OpenClaw compatibility
openclaw plugins install graph-memory
openclaw plugins enable graph-memory
openclaw gateway restart
Activate the Context Engine slot in ~/.openclaw/openclaw.json:
{
"plugins": {
"slots": { "contextEngine": "graph-memory" },
"entries": { "graph-memory": { "enabled": true } }
}
}
Earlier OpenClaw seven-turn token comparison
Graph Memory Pro
The repository also contains an experimental read-only DSH Pro Lite Host + Client plugin backed by Community SQLite. The 2D/3D graph workbench, split view, and controlled drag-to-context remain planned. See dsh-pro/README_CN.md.
Verification and limits
Current release 2.0.4 passes 491/498 automated tests (7 skipped), both TypeScript builds, npm package verification, and a live compatibility harness against the installed DSH 0.1.7-rc.2 runtime (session format V4 admission over every durable surface write).
-
DSH 0.1.7 compatibility: durable surface writes use the producer-owned source kind plugin:graph-memory; the retired V3 wrapper { kind: "plugin", plugin: "graph-memory" } is rejected by V4 persistence admission and fails the whole turn once rolling compaction or tool-trace projection triggers (#113).
-
DSH 0.2 line host support (0.2.0-rc.2 and 0.2.1-rc.2): package.json declares engines.dsh >=0.1.7-rc.2 <0.3.0 (per app-boot, once engines.dsh is present it alone gates DSH compatibility and peer ranges are ignored), and the @deepseek-ai/dsh-typert-protocol peer is capped at the same <0.3.0 (>=0.1.0-rc.5 <0.3.0). The 0.2 line ships the same core host deps Graph Memory is validated against (cordis ~4.0.4/~4.0.5, schemastery ~3.18.4/~3.18.5) and session format stays V4 (the 0.2.1 "V4.1" refers to the DeepSeek-V4.1-Flash-1017 model, not the message format). Surface checks against 0.2.1-rc.2 found: dsh-client-ui-primitives still exports the SettingsForm/SegmentedControl/Switch the client uses; dsh-typert-protocol@0.2.1-rc.2 still exports Remote/TypertRemoteService; and none of the 0.2.1 breaking changes (./invariant export removal, composer stats split into activity/usage, webRuntime→webStartup rename, Agent Team inbox reroute, both tool display-mode removal, azure-openai-responses→azure rename) touch this plugin's surface. 0.2.0-rc.2 predates every one of those breaking changes.
-
Structured extraction still depends on model contract compliance: the latest run succeeded 20/20 times; any future failure stays quarantined and never blocks the foreground conversation.
-
Recall is bounded by configurable Top-K. Focused probes succeeded; one broad multi-topic query can require a larger Top-K or separate questions.
-
The published run is an engineering workload, not a universal LoCoMo/LongMemEval score.
-
The design, source-code map, and porting sequence for the summary + SPO navigation + exact-Q/A upgrade are documented in the .
Reproduce it from benchmarks/dsh-context-takeover/. Raw conversations, provider responses, local paths, and credentials are excluded.
Development
npm install
npm test
npm run build
npm run verify:package
MIT © 2026 adoresever · Asset and trademark notes
