dsh-thinking-language-zh
ayanJava111
deepseek harness思考过程中文插件
PROJECT TOPICS
PROJECT README
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.
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 |
| 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.
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.
"dsh": { "profile": { "bundles": ["@bryan-cmf/dsh-vector-memory"] } }
⚠️ This row publishes the
vectorMemoryservice — it must live in the host composition (process-global). If you mount it from an agent preset instead, wrap it in a group with anisolaterealm (see the DSH composition docs).
>= 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)| Key | Default | Meaning |
|---|---|---|
ttlDays |
90 |
Default memory TTL; 0 = forever |
maxResults |
10 |
Default mem_search result count |
pnpm install --store-dir .pnpm-store
pnpm build
MIT
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。