dsh-memory

dsh-memory

Cross-session semantic memory for DeepSeek Harness — installed into any profile, so your agent actually remembers you across sessions.

🌐 English · 简体中文

---

dsh-memory registers three tools for the web / headless profiles, so your agent can remember you across sessions:

• memory_add — persist a fact worth keeping long-term, auto-vectorized on the way in.
• memory_search — recall relevant facts by semantics (cosine similarity), with a keyword fallback.
• memory_update — correct or retire an existing memory (supersede / retract / restore / verify / edit / set / link); the superseded entry stays traceable instead of being deleted.
• /mem <question> — type this in the composer and the agent will call memory_search first (results render as a collapsible tool card), then answer from memory.

Under the hood: Zhipu embedding-3 (2048-dim) + SQLite (node:sqlite) + cosine similarity. Zero new infrastructure — no vector database, no sidecar containers. Hundreds to thousands of memories recall in milliseconds.

Why

Your agent forgets everything between sessions. dsh-memory gives it a durable, semantic recall layer without adding a new service to run. It solves the "who am I / what did we agree on" problem with a few hundred bytes of SQLite and one HTTP call per write.

Features

• Zero-config, never crashes — a missing Zhipu key, a missing database, a missing tools service, or a failing embedding call all degrade gracefully. This plugin will never take the DSH tree down (see Design guarantee).
• Semantic recall with keyword fallback — memory_search scores by embedding cosine similarity first, then weights keyword hits; when embeddings are unavailable it falls back to pure keyword (bigram) matching.
• 8s embedding timeout — on a bad or missing network it fails fast and falls back to keyword search instead of hanging the session.
• Correction chain, not silent merging — when a fact changes, pass replaces to memory_add: the old entry is stamped superseded + superseded_by and kept for traceability (it drops out of default search), while the search that would have returned it now surfaces the corrected conclusion — even when that conclusion only ranks 12th. Writing still runs one cosine pass over the library first, but it reports near-duplicates as a hint (replaced) and never silently merges two different facts.
• Annotate, never filter — scope, date and status are shown as information; they are never used as hard filters, and relevance is reported as high/medium/low instead of a threshold deciding "nothing found" (measured: genuine and absent answers overlap completely in cosine similarity). Entries over 800 chars are truncated by default, with the full length shown; full: true returns the whole thing.
• SQLite WAL + busy_timeout=5000 — applied when the DB opens; concurrent writers (e.g. DSH web + the Jarvis brain sharing one library) no longer hit SQLITE_BUSY.

Design notes, the measured evidence behind each choice, and the assumptions that were disproved by data (so nobody re-walks them) live in LIFECYCLE.md.

Installation

This plugin follows the official DSH bundle convention (dsh.bundle in package.json), so dsh plugin recognizes it and activates it as a configuration layer — not a plain dependency.

Install into the target profile (e.g. web):

dsh plugin --profile web add dsh-memory

After installing, restart DSH web for it to take effect (the bundle layer is only composed at startup):

systemctl restart dsh    # or restart however you run DSH

Local / pre-release install:

# Put this directory into the profile's node_modules and append "dsh-memory"
# to dsh.profile.bundles in package.json, then restart.

Configuration

Everything is optional — skip it all and the plugin still works (falling back to keyword search):

forceMemoryWords example (set in your profile's cordis.patch.yml, not in the public repo):

- id: dsh-memory
  config:
    forceMemoryWords:
      - 记住
      - 重要
      - 珍贵
      - 特别
      - 务必
      - 一定
      - 必须

Zhipu credentials are read from environment variables first, then from the file at zhipuEnvPath:

ZHIPU_API_KEY=your-zhipu-key
ZHIPU_BASE_URL=https://open.bigmodel.cn/api/paas/v4/

Get a Zhipu key for free at open.bigmodel.cn (embedding-3 is billed per use).

Design guarantee

This plugin was born from a real incident: a misconfigured plugin sent a DSH web profile into a crash-loop. So dsh-memory is deliberately defensive:

1. apply() is wrapped in an overall try/catch — no error is ever thrown up into the DSH tree.
2. inject only declares tools (a service dsh-base always provides). It does not depend on dsh-llm, agents, or http.
3. A missing key / database / service, or a failed embedding call → it logs and degrades, never interrupting startup.
4. Errors raised inside a tool's execute() are caught and returned as ordinary results, never bubbling up into a session exception.

Tools

• memory_add — write a fact worth remembering; vectorized automatically.
• memory_search — semantically recall the most relevant facts for a question.
• memory_update — supersede / retract / restore / verify / edit / set / link an existing memory (superseded entries stay traceable).
• /mem — a manual override: "search memory first, then answer." If you're not sure the agent will recall on its own, type /mem <question> and it will run a semantic search, inject the results, and answer from them.

Storage schema

memories(id, text, category, source, created_at, updated_at,
         status, kind, kind_source, scope, scope_source, superseded_by, last_verified_at)
memory_embeddings(memory_id, dim, vector)   -- 2048-dim JSON array
memory_links(a, b, relation, created_at)    -- links between related-but-different entries
memory_meta(key, value)                     -- schema_version

License

MIT