@mindmemos/deepseek-harness-plugin

A DeepSeek Harness (dsh) plugin that wires MindMemOS long-term memory into the harness. It:

• recalls memories before each turn by running mindmemos memory search and injecting the hits as model context, and
• stores each completed turn by running mindmemos memory add.

It is a thin shell-out layer over the mindmemos CLI — the same integration the OpenClaw plugin provides for a different host.

Prerequisite: the mindmemos CLI must be installed and authenticated first. Installing this plugin alone does nothing — it spawns the CLI for every operation.

Install the CLI

# whichever installer you use for the MindMemOS CLI; e.g.
pip install mindmemos-sdk
mindmemos auth

Confirm it works from the shell dsh will run under:

mindmemos config show

Install the plugin

Install the published package in the project where dsh runs:

npm install @mindmemos/deepseek-harness-plugin

For local development before publishing, build from this repo instead (see Build) and register the built entry by file:// URL.

Build

cd plugins/deepseek-harness-plugin
npm install
npm run build        # emits dist/index.js + dist/index.d.ts

npm run typecheck runs the same compiler pass without emitting — use it to validate against the dsh types before wiring the plugin up.

Register

dsh composes plugins through layered cordis.patch.yml files. Add an insert entry to your profile patch ($DSH_HOME/cordis.patch.yml or the profile's cordis.patch.yml):

- insert:
    - id: mindmemos-memory
      name: '@mindmemos/deepseek-harness-plugin'   # once published
      config:
        userId: alice
        appId: deepseek-harness

For local testing before publishing, point name at the built entry (see cordis.patch.example.yml for the exact form):

- insert:
    - id: mindmemos-memory
      name: 'file:///C:/…/plugins/deepseek-harness-plugin/dist/index.js'
      config:
        userId: alice

id is stable and unique; the plugin's cordis name is mindmemos-memory. To disable the plugin, set disabled: true on the entry rather than removing the row.

Configure

Option	Default	Meaning
cli	mindmemos	Executable used to invoke the CLI — a name on dsh's PATH or an absolute path. It is not a shell command, so a wrapper like uv run mindmemos will not work; point it at the real executable instead.
topK	5	Number of memories injected per turn.
addMode	async	sync blocks until extraction finishes; async enqueues and returns. In async mode only CLI-level failures are visible to the plugin.
userId	(none)	Scopes both search and add to one user. Omit for project-wide search; add then inherits the CLI's default user.
appId	deepseek-harness	Application scope attached to every search and add.
sessionId	(none)	Override the harness session id used as the CLI session scope.
minQueryLength	2	Skip recall for prompts shorter than this many characters.
maxConversationMessages	80	Cap on how many trailing messages are persisted per turn.

How it works

• Recall hooks agent/pre-step (step 1 only) as a prepended waterfall listener. It extracts the real human prompt, runs mindmemos memory search, and appends one createUserMessage carrying the hits under a `` banner. The injected message is stamped source.kind === "plugin" so the store step can tell it apart from real input.
• Store hooks the session/event firehose and reacts to turn/end with reason.kind === "completed". It walks that turn's log back to its turn/start, collecting the surface user/assistant/tool messages, and runs mindmemos memory add --messages-json-file -. Plugin-injected context (source.kind !== "user") is excluded so recalled memories are not re-stored.

Verify

1. Types build cleanly

   npm run typecheck
   

2. Recall fires

   First store a fact out of band, then ask about it in dsh:

   mindmemos memory add --messages-json-file - --json --user-id alice --app-id deepseek-harness <<'JSON'
   [{"role":"user","content":"My favorite color is teal.","timestamp":0}]
   JSON
   

   In dsh, ask "what is my favorite color?". The plugin log should show:

   [mindmemos-memory] recall hit 1 memories, injected N chars
   

   and the model should answer with the stored fact.

3. Store writes

   Complete any turn in dsh, then confirm the log shows:

   [mindmemos-memory] stored N message(s) from turn 1 (session_id=…)
   

   and that the memory is searchable afterwards:

   mindmemos memory search "…" --json --user-id alice
   

Troubleshooting

• ENOENT / command not found — dsh's PATH may not include the CLI (common for GUI-launched dsh). Set cli to the executable's absolute path (which mindmemos, or uv run which mindmemos inside a uv project). cli is a single executable path, not a shell command, so a wrapper like uv run mindmemos will fail with ENOENT.
• Recall never fires — check minQueryLength and that cli resolves (mindmemos config show from dsh's environment).
• Recall works but nothing is stored — the store only runs when a turn completes (reason.kind === "completed"); aborted or errored turns are skipped by design. Check the log after a turn that ends cleanly.
• CLI errors surface as [mindmemos-memory] memory search failed: … / [mindmemos-memory] memory add failed: … warnings, including the CLI's stderr and exit code.