메뉴
Use ChatMem for repository recall, continuation, handoff, migration, startup rules, wiki/context, local history search, and durable memory. 当用户问“记得吗、之前讨论过、回忆、继续做、迁移、交接、项目历史、本地历史、启动规则、记忆”等问题时使用；先查 ChatMem MCP，不要让用户重新描述已有对话。
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/Rimagination/ChatMem --skill chatmem명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SKILL.md 표시 중
SOC 직업 분류 기준
| name | chatmem |
| description | Use ChatMem for repository recall, continuation, handoff, migration, startup rules, wiki/context, local history search, and durable memory. 当用户问“记得吗、之前讨论过、回忆、继续做、迁移、交接、项目历史、本地历史、启动规则、记忆”等问题时使用；先查 ChatMem MCP，不要让用户重新描述已有对话。 |
ChatMem is an MCP-first local-history, startup-rule, and continuation layer. Indexed local history is the evidence layer; approved startup rules are compact durable guidance for new agent sessions; generated wiki pages are readable projections rebuilt from approved rules and episodes. This skill is only the operating guide: use it to decide when and how to call the ChatMem MCP server.
Do not ask the user to paste full historical transcripts when ChatMem can provide repo memory, checkpoints, handoffs, or targeted history.
Do not assume ChatMem appears as an @chatmem mention in the chat UI. For agents, ChatMem should normally be reached through MCP tools; the desktop app is the human-facing conversation and memory control surface.
Use this skill when the user:
Do not use it for general web search, unrelated memory, or one-off notes that will not help future work in the same repo.
要我先用 ChatMem 低成本回忆一下这个项目吗？我会只加载启动规则、最近交接和少量相关历史，不展开完整对话。recall_project_work first with limit=3. This is the default high-level recall tool: it returns status, compact answer, de-duplicated evidence, diagnostics, and next actions, and it does not run foreground import or scan.get_project_context when startup rules, handoff metadata, pending candidates, and compact history are all needed together. Use intent="startup" for compact startup rules, intent="recall" when the user asks whether something was discussed before, and intent="continue_work" when resuming.limit=3 unless the user asks for a broad review. Read approved rules, recent handoff metadata, diagnostics, and the top relevant history summaries first; do not load or paste raw transcripts.recall_project_work, get_project_context, or get_repo_memory_health reports unmatched project roots that clearly belong to the same repo, use merge_repo_alias, then run scan_repo_conversations as maintenance and retry recall.recall_project_work is unavailable, fall back to get_project_context and then call search_repo_history for specific gaps: prior decisions, commands, key files, errors, earlier attempts, or recall questions. When a hit includes conversation_id, use read_history_conversation only after the user asks you to read/expand that found conversation.list_repo_wiki_pages or rebuild_repo_wiki when the user asks for a readable project wiki, commands, gotchas, or recent-work pages. Treat those pages as generated projections, not editable source material.list_entity_graph when a task depends on related systems, symbols, protocols, agents, or release concepts.list_memory_conflicts before writing or updating a surprising startup rule, especially when it negates an existing command or convention.create_memory_candidate with concise text and evidence.propose_memory_merge or the relevant MCP write path to prepare an agent-authored rewrite. Do not silently overwrite approved rules; ask the user first when the evidence or conflict is uncertain.build_handoff_packet instead of asking the user to copy the full conversation.Use this ladder whenever the user agrees to project recall, asks "do you remember...", or needs to continue prior work:
recall_project_work with the exact topic and limit=3.found, summarize the returned answer, source agent/conversation, and evidence label. Say that this is indexed local history, not an approved startup rule. Ask whether the user wants you to read the most relevant conversation with read_history_conversation before increasing the limit or requesting broader history.partial, try one narrower query with exact file names, feature names, or conversation titles. If the miss appears caused by repo-root drift, use merge_repo_alias or retry with the likely related repo root.miss and diagnostics show no indexed documents/chunks, say the indexed coverage is missing and run import_all_local_history or scan_repo_conversations only as a background maintenance step before concluding history is absent.get_project_context(intent="recall", limit=3) only when you also need approved startup rules, recent handoff, or pending candidate context.limit=10 or more only after the user asks for a broader review or after the first compact pass proves the topic is relevant.If recall_project_work.evidence, get_project_context.relevant_history, or search_repo_history.matches returns any plausible match:
source_agent or the prefix in conversation_id), conversation title, updated date, and evidence excerpt when available.我在 Codex 和 Claude 的本地历史里找到了关于 Pmodel 的对话证据。要我读取最相关的那段对话，还是先列出命中的对话让你选？read_history_conversation with the returned conversation_id, the evidence message_id when present, the original query, and a compact limit such as 8-12.read_history_conversation is not available in the current MCP tool list, say the installed ChatMem MCP is older and continue with the existing search evidence: offer to list the matched conversations, run search_repo_history with a slightly larger limit, or ask the user to open the conversation in ChatMem. Do not ask the user to redescribe the topic.The ChatMem desktop app may copy a short prompt like:
Use ChatMem to load memory for this project:
repo: D:\VSP
conversation: 019dab66-4427-7d93-a7dc-6cb90a1a4a74
Prefer ChatMem MCP over pasted history.
When you see this, do not treat it as context by itself. Treat it as an instruction to call ChatMem MCP for the repo and optional conversation id.
Language rule:
跨 agent 记忆依赖 repo_root 归一化；继续使用 canonical_repo_root 匹配 .git 根目录。Good candidates are durable, repo-scoped, and useful as startup rules:
Bad candidates:
The desktop settings screen owns WebDAV credentials and cloud upload. MCP tools do not silently write to the user's cloud storage. When a user expects files to appear in their netdisk, the app must run the explicit WebDAV sync action, which creates the remote chatmem folder and uploads JSON conversation snapshots plus a manifest.
Search history is indexed as FTS plus vectors. Local chatmem-local-hash-v1 vectors are always kept as a fallback. When CHATMEM_EMBEDDING_PROVIDER=openai-compatible and the matching base URL/model/dimensions/API key environment variables are present, search can use real provider embeddings stored side-by-side with the local fallback. Call rebuild_repo_embeddings after changing embedding provider settings. Treat vector hits as retrieval candidates, not final truth: prefer approved startup rules and evidence refs for durable facts, and verify stale or surprising matches before acting.
For recall questions, never answer from get_repo_memory alone. If approved startup rules do not contain the answer, search history and clearly label matches as history evidence rather than approved rules. If history evidence exists, offer to read/expand the found conversation before asking the user to explain the topic. If exact repo search misses, broaden to the repo family: parent roots and child roots can contain the real conversation.
import_all_local_history is the broad bootstrap tool. Use it sparingly: first install, major source-path changes, or suspicious recall misses.scan_repo_conversations is the repo-scoped maintenance tool. It links conversations that match the repo and returns unmatched project roots for diagnosis.merge_repo_alias is the repair tool for path drift: old cwd, file cwd, renamed folder, generated Codex project path, or related subproject that should belong to the current repo.scan_repo_conversations again before answering recall questions.ChatMem may auto-create pending startup-rule candidates only from explicit durable-memory markers such as "Remember:", "Rule:", "Gotcha:", "记住：", or "规则：". Bare imperative lines such as "Always ..." or "Do not ..." are not enough because they often come from agent task instructions. Auto-extracted candidates are suggestions only and are not required for local-history search. The desktop app is for inspecting or deleting them; promotion into startup rules belongs to the active agent through MCP or skill-driven workflow after checking evidence and conflicts.
ChatMem does not need an internal LLM API to rewrite startup rules. The active agent can draft or apply a better rule update through MCP or skill workflow after reading the candidate, approved rule, conflicts, and evidence. The desktop inbox remains a view/delete surface, not an approval queue. If the rewrite is uncertain, ask the user before writing or submitting a confident proposal.
If ChatMem MCP tools are not available:
Avoid telling the user to paste an entire historical conversation unless there is no other workable path.
When you use ChatMem, be brief:
Do not expose internal memory mechanics unless the user asks. The user experience should feel like "continue this project", not "manage a memory database".