写入一条记忆短文本（事实/笔记/摘要），可选标签与归属项目；返回 {"id": 记录ID}。 内容为空串或纯空白时返回 {"error": "INVALID_ARGUMENT", "message": 原因}。 v2（2026-08-30）：身份由环境承载——client 自动从 MCP 握手 clientInfo 识别 （TraeWork / Claude Code / Cursor），project 为项目/任务归属（连接配置声明， 可空=该客户端默认桶）；记录主键由服务端生成。共享知识（add_document/add_webpage） 所有客户端可见。 English: Write a memory short text (fact/note/summary) with optional tags and project; returns {"id": record_id}. Returns {"error": "INVALID_ARGUMENT", "message": reason} when content is empty or blank. v2: identity comes from the environment — client is auto-detected from the MCP clientInfo handshake (TraeWork / Claude Code / Cursor), project is the project/task bucket (declared in the connection config; empty = this client's default bucket). The record primary key is generated server-side. Shared knowledge (add_document/add_webpage) is visible to all clients.

混合检索记忆与知识（向量语义 + BM25 关键词，RRF 融合）； 返回命中列表，每项含 id/content/score/type/source。 v2：个人记忆（memory）只返回归属当前 (client, project) 的； 共享知识（doc/web chunk）所有客户端可见。top_k 小于 1 时返回 {"error": "INVALID_ARGUMENT", "message": 原因}。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 English: Hybrid retrieval over memories and knowledge (vector semantics + BM25 keywords, RRF-fused); returns a hit list, each item having id/content/score/type/source. v2: memory records only return those owned by the current (client, project); shared knowledge (doc/web chunks) is visible to all clients. Returns {"error": "INVALID_ARGUMENT", "message": reason} when top_k is less than 1. client: source client (optional; auto-detected from clientInfo when omitted).

按 ID 读取单条记忆完整内容；记录不存在返回 {"error": "NOT_FOUND"}； 非 (client, project) 归属的 memory 返回 {"error": "FORBIDDEN"}；共享知识可读。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 project：项目归属（可选，缺省=该客户端默认桶）。 English: Read the full content of a single memory by ID; returns {"error": "NOT_FOUND"} when absent; a memory not owned by the (client, project) returns {"error": "FORBIDDEN"}; shared knowledge is readable.

按 ID 更新记忆内容（变更后自动重新嵌入）；记录不存在返回 {"error": "NOT_FOUND"}； 非 (client, project) 归属返回 {"error": "FORBIDDEN"}； 内容为空串或纯空白时返回 {"error": "INVALID_ARGUMENT", "message": 原因}。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 English: Update a memory's content by ID (auto re-embed on change); returns {"error": "NOT_FOUND"} when absent; non-(client, project) owners get {"error": "FORBIDDEN"}; empty content returns {"error": "INVALID_ARGUMENT", "message": reason}.

按 ID 删除一条记忆；成功返回 {"ok": true}，记录不存在返回 {"error": "NOT_FOUND"}； 非 (client, project) 归属返回 {"error": "FORBIDDEN"}。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 English: Delete a memory by ID; returns {"ok": true} on success and {"error": "NOT_FOUND"} when it does not exist; non-(client, project) owners get {"error": "FORBIDDEN"}.

导入本地文档（PDF/DOCX/MD/TXT 及 Office 格式）切分入库； 返回 {"source": 文件名, "chunks": 块数}；文件不存在或格式不支持时 返回 {"error": "FILE_NOT_FOUND" | "UNSUPPORTED_FORMAT", "message": 原因}。 project/client：仅用于审计归类；文档 chunk 为共享知识，所有客户端可检索。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 English: Import a local document (PDF/DOCX/MD/TXT and Office formats), split and ingest it; returns {"source": filename, "chunks": count}; when the file is missing or the format is unsupported returns {"error": "FILE_NOT_FOUND" | "UNSUPPORTED_FORMAT", "message": reason}. project/client: audit bucketing only; document chunks are shared knowledge, searchable by all clients.

抓取网页正文并切分入库；返回 {"source": url, "chunks": 块数}； 抓取/正文提取失败时返回 {"error": "WEB_FETCH_FAILED", "message": 原因}。 project/client：仅用于审计归类；web chunk 为共享知识，所有客户端可检索。 client：来源客户端（可选，缺省从 clientInfo 自动识别）。 English: Fetch a webpage body and ingest it after splitting; returns {"source": url, "chunks": count}; on fetch/body-extraction failure returns {"error": "WEB_FETCH_FAILED", "message": reason}. project/client: audit bucketing only; web chunks are shared knowledge, searchable by all clients.

基于知识库的 RAG 问答（检索 → 上下文拼装 → 护栏生成），返回 answer 与 sources； v2：检索按当前 (client, project) 隔离 memory（doc/web 共享），并写 ask 存取审计； client：来源客户端（可选，缺省从 clientInfo 自动识别）； LLM 不可用时返回 {"error": "LLM_DISABLED", "message": 配置指引}。 English: Knowledge-base RAG Q&A (retrieve → build context → guarded generation), returning answer and sources; v2: retrieval isolates memory by the current (client, project) (doc/web shared); an ask access-audit is emitted; client: source client (optional; auto-detected from clientInfo when omitted); returns {"error": "LLM_DISABLED", "message": setup guidance} when the LLM is unavailable.