dsh-memory
Persistent memory system for DeepSeek Harness (dsh). Provides multi-channel recall, confirmed writes, an encrypted secrets vault, session transcription, and maintenance tooling as a dsh plugin.
记忆系统插件：RRF 多路召回 / 确认制写入 / Fernet 加密库 / 会话转录 / 维护工具。
Features
- RRF 五路召回 (five-way RRF recall) — fuse evidence from L1 rules (MEMORY.md / USER.md), L2 session FTS5, L3 semantic (optional Honcho/MCP), L4 wiki, and L5 time-window sources into a single ranked evidence ledger with per-item trust labels.
- 确认制写入 (confirm-before-write) — nothing mutates your memory directly:
memory_suggest enqueues a proposal, memory_confirm / memory_reject decide it. Budget checks, sensitive-pattern scanning, and automatic trigger: tags guard what lands in L1.
- Fernet 加密库 (encrypted vault) — secrets live in a Fernet-authenticated encrypted store (
VAULT.md.enc); plaintext appears only as → vault:<ref> references.
- 会话转录 (session transcription) — incrementally indexes dsh session logs into an FTS5 store and extracts correction signals for skill distillation.
- 维护工具 (maintenance) — health checks, git backups, hit auditing, and an asset de-duplication checker.
Tools
| Tool | Description |
|---|
memory_recall | Five-way RRF recall with source labels and trust levels; date= for time-window queries, deep=true for deeper recall |
memory_graph | Entity graph walk (curated relation edges + automatic co-occurrence edges) |
memory_suggest | Enqueue a memory write proposal (track: memory / user / skill; lifecycle: rule / stable / situational / record) |
memory_confirm | Accept a pending proposal into MEMORY.md / USER.md / skill |
memory_reject | Reject a pending proposal |
memory_vault_get | Read an encrypted vault entry |
memory_vault_set | Write an encrypted vault entry |
memory_harvest | Transcribe dsh session logs into the L2 FTS5 index and extract correction signals |
memory_health | One-line-per-layer health report (L1 budgets, vault decryptability, L2 index, graph freshness, pending queue) |
memory_backup | Git commit + push of the memory directory |
memory_notify | Push a message to a Feishu/Lark chat (requires credentials, see below) |
check_assets | De-duplication check before creating new scripts/skills/plugins/automations |
Installation
dsh plugin --profile web add github:boomzikazita/dsh-memory
The plugin registers its tools and hooks (primer recall injection, write guard, turn-end transcription) automatically. Python scripts under scripts/ are executed on demand by the Node host.
Configuration
The bundle patch (cordis.patch.yml) mounts the plugin with defaults; override per profile, e.g.:
- insert:
- id: dsh-memory
name: 'dsh-memory'
config:
memoryDir: '~/.dsh/memories'
# scriptsDir: '~/.dsh/plugins/dsh-memory/scripts'
# wikiRoot: '~/wiki' # env DSH_WIKI_ROOT also works
# guardBlock: false
# primerBudget: 800
# primerTopK: 3
Key paths (all overridable):
| Setting | Default | Notes |
|---|
memoryDir | ~/.dsh/memories | L1 files, FTS5 indexes, vault, outbox |
scriptsDir | plugin scripts/ dir | Python helper scripts |
wikiRoot / DSH_WIKI_ROOT | ~/wiki | Obsidian/wiki root used by the L4 channel (raw/, _meta/, entities/ live under it) |
Dependencies
- Node side:
@deepseek-ai/dsh-tools and @deepseek-ai/dsh-llm (declared as peer dependencies, >=0.1.0-rc.6).
- Python side (optional, for recall/graph/vault/harvest features): Python 3 with the standard library; SQLite 3. The vault additionally requires
cryptography (Fernet). If Python is missing, tools degrade gracefully with an error message.
- Feishu/Lark push (
memory_notify / scripts/feishu_push.py): set LARK_APP_ID / LARK_APP_SECRET (env or ~/.env) and optionally LARK_DEFAULT_CHAT for the default chat target.
- L3 semantic recall (optional): set
DSH_HONCHO_MCP_URL / DSH_HONCHO_MCP_TOKEN to an MCP endpoint exposing honcho_search. Unset → L3 silently skipped.
License
MIT — see LICENSE. Copyright (c) 2026 boomzikazita.
