dsh-memvault

Long-term memory for DeepSeek Harness.
blocks in the prompt · turns extracted into one SQLite file · two tools · one panel

English | 中文

---

MCP is pull-only, so "memory is already loaded" can never come from a server. This plugin does it inside DSH — and since 0.9.x it owns the store too: no Python, no second process.

Half	What	How
Inject	core blocks in the system prompt, every step	ctx.systemPrompt.context()
Extract	finished turns → one window → one LLM pass	session/event → window → add/supersede/none
Recall	memory_search / memory_add	FTS5 trigram + vectors, RRF, answerability guard
Show	overview / memories / structure, block editing, replay	host routes + conversation tab / settings page

✨ Features

🗄️ One store	prompt, extraction, tools and panel read ~/.dsh/memvault/memvault.db
🧠 No tool call needed	blocks ride the runtime context, TTL-cached
⏱️ Off the critical path	queued windows, one call per N turns / idle / cap
🔎 Hybrid recall	substring truth (trigram) + paraphrase (vectors), fused
🛠️ Two tools	scoped by user_id / agent_id / run_id
🎛️ A real panel	diagnostics, vector backfill, provenance, relation graph
🪶 Zero deps	node:sqlite + hand-written client bundle

🚀 Install

dsh plugin --profile desktop add dsh-memvault

Self-owned install links a checkout in the profile's package.json (dsh-memvault: link:…) plus a junction under node_modules — see docs/DEVLOG.md. A new bundle needs a DSH restart; after editing the client run npm run build.

⚙️ Configuration

Field	Default
scopes	[{user,lenovo},{agent,claude-code-memory}]
labels / maxChars / refreshMs	[] / 4000 / 30000
extract.everyNTurns / idleMs / windowTurns	3 / 20000 / 8
extract.minTranscriptChars / maxInputChars	40 / 6000
extract.projectDir / env	checkout path / {} (where .env is read)
panel.writes	true

dbPath and extract.pythonPath are accepted but unused. The pipeline reads the usual .env keys (OPENAI_*, MEMVAULT_EXTRACTOR=rule|llm, MEMVAULT_EMBEDDER=local|openai, MEMVAULT_ADJUDICATOR=off).

🗄️ Store & migration

~/.dsh/memvault/memvault.db (override DSH_MEMVAULT_CORE_DB). From the Python store, once:

memvault export --from data/memvault.db --out export.json
node scripts/import-core.mjs export.json ~/.dsh/memvault/memvault.db --apply
node scripts/backfill-core.mjs ~/.dsh/memvault/memvault.db --project-dir <checkout> --apply

🏗️ How it works

flowchart LR
  T[turn/end] --> W[window] --> X[extractor] --> D[decide] --> S[(store)]
  S --> P[system prompt]
  S --> M[tools]
  S --> U[panel]

🧪 Verification

npm test (6 smokes + 48 unit). Live: /memvault/api/status shows the last five outcomes ([rule]|[llm] ok added=N); replay via POST /memvault/api/replay; recall via memory_search.

🧠 Decisions

• One store, not two — no "tool saw it, panel didn't".
• Supersede, never delete — history keeps the audit.
• Only user text is extracted — short windows are skipped (docs/DEVLOG.md §35).

⚠️ Limits

• Single machine, single store: no sync, no encryption.
• Extraction quality depends on the extractor (rule is deterministic but narrow; LLM depends on the endpoint).

📄 License

MIT