@neatmem/dsh-neatmem

DeepSeek Harness (dsh) plugin for using NeatMem as a long-term memory backend.

The plugin is a thin REST client: it handles recall injection, turn capture, and memory tools, while the NeatMem server does extraction, deduplication, batching, and recall filtering. No native dependencies — nothing to compile or approve at install time.

Quick Start

Prerequisites:

• A running NeatMem server (pip install neatmem && neatmem serve, listens on http://localhost:8790 by default)
• A dsh installation with a configured model route (verified against dsh 0.1.5-rc.2)

dsh plugin --profile <name> add @neatmem/dsh-neatmem

Restart the profile to load the plugin. Verify the bundle composed:

dsh --profile <name> --dump-config   # shows a `neatmem-dsh` row

The plugin works with zero configuration — defaults are baseUrl=http://localhost:8790, userId=default, recall/capture/tools all enabled. At startup it pings the server once; if unreachable it warns and degrades (chats keep working, no recall/capture) until the server is back.

Configuration

Override any field per profile in $DSH_HOME/profiles/<name>/cordis.patch.yml:

- id: neatmem-dsh
  config:
    baseUrl: http://localhost:8790
    userId: myname
    topK: 5

Field	Default	Purpose
enabled	true	Master switch
baseUrl	http://localhost:8790	NeatMem server URL
apiKey	''	Optional Token auth header
userId	default	Memory namespace (user_id)
agentId	''	Optional agent-preset scope (agent_id)
recallEnabled / captureEnabled / toolsEnabled	true	Feature switches
recallTimeoutMs	3000	Auto-recall budget (hard-capped at 3s, fail-open)
topK	5	Max memories injected per turn
searchThreshold	0.3	Minimum relevance score
contextMaxChars / toolResultMaxChars	6000 / 1200	Bounded rendering

Invalid configuration fails at load time; an unreachable server only downgrades memory, never the chat.

How It Works

Auto-recall — on each direct-user turn's first step, the plugin runs one bounded search (rerank off, fast path) and injects hits as a source-labelled user message (plugin/neatmem-dsh/recall), reconstructable from the session log. Plugin/tool messages never trigger recall.

Auto-capture — each finished turn (user + assistant text) is forwarded via POST /v1/messages/add/ (store-first; the server's scheduler batches extraction). On session close, a bounded flush (/v1/messages/flush/) forces extraction of anything pending. Capture runs on a serial per-session queue and never blocks a turn.

Memory tools — the agent can call memory_search (exposes rerank for accuracy), memory_get, memory_list, memory_update, memory_delete. There is intentionally no memory_add: capture is automatic.

Scoping: run_id maps to the dsh session id (enables sessionScope in tools), user_id is your configured namespace.

Development (from source)

cd /path/to/NeatMem/dsh
pnpm install
pnpm run build

cd /path/to/deepseek-harness
pnpm dsh plugin --profile <name> add /path/to/NeatMem/dsh

Rebuild after changing TypeScript source, then restart the profile.

Publishing

npm publish (prepublishOnly rebuilds dist; the tarball ships only dist + cordis.patch.yml).