ReMe/integrations/dsh

ReMe Memory for DeepSeek Harness

This DSH bundle follows the same separation used by QwenPaw's embedded ReMe integration:

• the main agent receives durable memory guidance;
• the model can call reme_search explicitly;
• completed user turns are submitted to auto_memory in the background;
• auto_dream runs as an independent daily maintenance task.

The ReMe HTTP service remains the owner of workspace files, indexes, memory extraction, and dream consolidation. The plugin does not copy or rewrite those files.

Requirements

• DeepSeek Harness 0.1.0-rc.7 or later compatible 0.1.x release
• Node.js ^22.19.0 or >=24
• A running ReMe HTTP service with the search, auto_memory, and auto_dream jobs enabled

Start ReMe against the workspace that should own the agent's memory:

reme start workspace_dir=/absolute/path/to/workspace

Install the published bundle into a DSH profile:

dsh plugin --profile default add @agentscope-ai/reme-dsh-memory

For a source checkout, build a package tarball first. A direct local-directory install only creates a link and does not run this bundle's build:

cd integrations/dsh
npm ci
bundle=$(npm pack)
dsh plugin --profile default add "./$bundle"

Configuration

The default endpoint is http://127.0.0.1:2333. Set REME_URL, or use the existing REME_HOST and REME_PORT variables. Bundle configuration can be added to cordis.patch.yml:

- insert:
    - id: reme-memory
      name: '@deepseek-ai/cordis-plugin-group'
      group: true
      isolate:
        remeMemory: true
      config:
        - id: reme-memory-runtime
          name: '@agentscope-ai/reme-dsh-memory'
          config:
            endpoint: http://127.0.0.1:2333
            language: zh
            timezone: Asia/Shanghai
            autoMemoryInterval: 5
            autoDreamEnabled: true
            dreamCron: '0 23 * * *'

Option	Default	Meaning
endpoint	http://127.0.0.1:2333	ReMe HTTP service URL
language	en	Memory guidance language: en or zh
autoMemoryEnabled	true	Capture completed main-agent turns
autoMemoryInterval	5	Submit after this many completed user turns
autoDreamEnabled	true	Enable daily dream maintenance
dreamCron	0 23 * * *	Daily schedule in the DSH process's local timezone
rootAgentsOnly	true	Keep prompt injection and capture out of subagents
requestTimeoutMs	10000	Search request timeout
backgroundTimeoutMs	3600000	Auto-memory and auto-dream timeout
shutdownTimeoutMs	5000	Maximum best-effort flush time while a session/plugin closes
timezone	Asia/Shanghai	IANA timezone used to split daily batches; must match the ReMe workspace

dreamCron intentionally accepts only the daily form <minute> <hour> * * *. This keeps the bundle dependency-free and makes the maintenance schedule explicit.

Runtime behavior

Memory guidance is injected as a source-attributed DSH user message rather than a system-prompt fragment. DSH presets may declare a complete persona and replace other system prompt contributions; a durable plugin message remains visible, replayable, and eligible for normal compaction.

Only direct human user messages and assembled assistant messages are sent to auto_memory. Plugin context and tool results are excluded so recalled text cannot be stored again as if the user had said it. DSH event sequence numbers are used to create stable ReMe message IDs, and DSH session IDs are mapped to fixed-length hashed ReMe session IDs.

Auto-memory calls are serialized per session and are never awaited by the model turn. If a request fails, its turns are put back into the in-process queue and retried after later activity. Turns are split into separate requests when their workspace dates differ. Session disposal makes one final best-effort attempt, bounded by shutdownTimeoutMs; it cancels outstanding HTTP work and retains any unconfirmed turns for a later plugin-shutdown retry. The current first version does not persist that retry queue across a DSH process crash or a failed final plugin shutdown.

Testing

cd integrations/dsh
npm ci
npm run typecheck
npm test
npm pack --dry-run

The plugin is authored in TypeScript under src/. npm run build emits ESM JavaScript, declarations, and source maps to the ignored dist/ directory. The prepare lifecycle builds the plugin when creating the npm tarball; install that tarball rather than the source directory. The tarball contains only the compiled dist/ output, this README, and the DSH bundle patch.

Publishing

Publishing is intentionally manual. Update package.json and package-lock.json to the release version, merge that change, then run the Publish ReMe DSH integration to npm workflow with the same version and the desired npm tag. The repository must provide an NPM_TOKEN Actions secret with publish access to the @agentscope-ai scope. The workflow type-checks, tests, packs, uploads the exact tarball as an artifact, rejects an already published version, and publishes that tarball with npm provenance.