@mem9/dsh-plugin

Persistent Mem9 memory for DeepSeek Harness. This package is an installable DSH/Cordis bundle. It adds automatic recall, background smart ingest, and five model-facing memory tools without changing the Harness agent loop or session format.

Install

export MEM9_API_KEY="<mem9-api-key>"
dsh plugin --profile <profile> add @mem9/dsh-plugin
dsh --profile <profile> --dump-config

Before the package is published to npm, install from a local checkout:

dsh plugin --profile <profile> add ./dsh-plugin

The bundle inserts one Cordis row named mem9. Override its complete config object in the profile's cordis.patch.yml when non-default settings are required.

Configuration

| Field | Default | Meaning | |---|---:|---| | apiUrl | https://api.mem9.ai | Hosted or self-hosted Mem9 origin | | apiKeyEnv | MEM9_API_KEY | Credential reference resolved through ctx.credentials for every operation | | agentId | deepseek-harness | X-Mnemo-Agent-Id request header and smart-ingest agent id | | defaultTimeoutMs | 8000 | Store/get/update/delete/ingest timeout | | searchTimeoutMs | 15000 | Search and automatic recall timeout | | includeSubagents | false | Give subagents tools, recall, and ingest when enabled | | recall.enabled | true | Search before the first model step of each user turn | | recall.minQueryChars | 5 | Skip shorter user text | | recall.limit | 10 | Maximum recalled memories | | recall.maxCharsPerMemory | 500 | Per-memory prompt bound | | ingest.enabled | true | Submit completed turns with mode: smart | | ingest.maxMessages | 20 | Maximum user/assistant messages per ingest | | ingest.maxBytes | 204800 | UTF-8 content budget per ingest |

Example profile override:

- id: mem9
  name: '@mem9/dsh-plugin'
  config:
    apiKeyEnv: MEM9_API_KEY
    apiUrl: https://api.mem9.ai
    includeSubagents: false
    recall:
      enabled: true
      minQueryChars: 5
      limit: 10
      maxCharsPerMemory: 500
    ingest:
      enabled: true
      maxMessages: 20
      maxBytes: 204800

Behavior

Recall runs at agent/pre-step for the first real user step. Results enter the ordinary session log as a plugin-sourced user/message with form: recall, so the model input is replayable without this package installed. Recalled text is explicitly labeled as untrusted historical context.

Smart ingest starts after a completed turn/end, selects only real user and assistant text, and never blocks turn completion. Work is serialized per session and drained with the plugin during shutdown. Automatic HTTP failures are fail-soft. A Mem9 runtime-quota denial adds at most one durable notice per session.

The model-facing tools are:

• memory_store
• memory_search
• memory_get
• memory_update
• memory_delete

Tools return structured JSON values. Expected API failures are returned as { "ok": false, ... }; quota errors preserve their safe action metadata. API keys are resolved per operation and never enter configuration output, session events, or tool errors.

By default, only top-level agents receive Mem9 registrations. Set includeSubagents: true to enable the same tools and automatic paths for subagents.

Development

pnpm install
pnpm test
pnpm typecheck
pnpm build
pnpm pack:check
pnpm test:e2e       # skips unless MEM9_API_KEY is set

The test suite exercises public Cordis, agent, session, tool, credential, HTTP, and package seams. Live E2E should use a dedicated test Space; it creates uniquely tagged data and deletes it in cleanup.