🧠 Vector Memory — DSH-Plugin

可持久化的 Agent 記憶核心——mem_save / mem_search / mem_health,跨 session、跨重啟,零外部依賴。

A DeepSeek Harness plugin that gives your agents durable, cross-session memory — the same mem_save / mem_search contract as the original Bryan-cmf/vector-memory MCP server, but natively on DSH.

Why a native rewrite (and what changed)

The 2026 memory landscape made self-hosting a Qdrant + BGE-m3 stack a commodity (Zep / Mem0 / Letta / LangMem all offer managed embeddings + graph memory). This package keeps the agent-facing contract and drops the infrastructure:

Aspect	Old (MCP server)	This plugin
Storage	Qdrant (Docker) + BGE-m3 (~2GB)	DSH storageDomain — durable JSON backend, zero setup
Retrieval	Vector similarity	v1: deterministic keyword scoring + recency; embeddings pluggable in v2
Interface	MCP stdio (Claude/Cursor/…)	Native tools (mem_save / mem_search / mem_health) + vectorMemory service
Persistence	Manual install script	Bundle row — it just works

Tools

Tool	Effect
mem_save	{ content, tags?, ttlDays? } → durable memory, returns id
mem_search	{ query, limit? } → ranked hits (score, content, tags, created)
mem_health	record count, expired count, domain status, TTL policy

Other plugins can inject: ['vectorMemory'] and call save / search / health directly.

Dashboard

The client half registers a 「記憶」view tab (order 30, right of chat / 軌跡 / 觀測) listing this session's saved/searched items live.

Install

"dsh": { "profile": { "bundles": ["@bryan-cmf/dsh-vector-memory"] } }

⚠️ This row publishes the vectorMemory service — it must live in the host composition (process-global). If you mount it from an agent preset instead, wrap it in a group with an isolate realm (see the DSH composition docs).

Requirements

• DSH >= 0.1.0-rc.6 — services: tools, storageDomain (domain layer + a storage backend, e.g. dsh-storage-json, are part of the shipped dsh-base bundle)

Configuration

Key	Default	Meaning
ttlDays	90	Default memory TTL; 0 = forever
maxResults	10	Default mem_search result count

Roadmap

• v2: pluggable embedding backends (hosted embedders or Qdrant REST)
• v2: MCP stdio dual-exit so the same store serves Claude Desktop / Cursor
• v2: memory curation (merge/dedupe/forget jobs)

Development

pnpm install --store-dir .pnpm-store
pnpm build

License

MIT