| name | claude-mem-mcp |
| description | Router and cheat-sheet for claude-mem MCP tools (prefix mcp__plugin_claude-mem_mcp-search__*). Use when users ask "did we solve this before?", "what did we do last session?", want to explore an unfamiliar codebase token-cheaply, or want to build a focused knowledge corpus from past work. Recognizes the 3-layer memory search pattern, AST-based smart_outline/unfold over Read, and the build→prime→query corpus lifecycle. |
claude-mem MCP router
Cheat-sheet for 11 tools under mcp__plugin_claude-mem_mcp-search__*, grouped into 3 workflows. For authoritative in-session refresher, call __IMPORTANT once per session.
When to use
- User references prior sessions ("did we fix this?", "last time").
- Exploring unfamiliar code — prefer AST tools over Read.
- Building durable Q&A brain over filtered history.
Tool inventory
| Family | Tools |
|---|
| Meta | __IMPORTANT |
| Memory search (A) | search, timeline, get_observations |
| Code exploration (B) | smart_search, smart_outline, smart_unfold |
| Knowledge agent (C) | build_corpus, list_corpora, prime_corpus, query_corpus, rebuild_corpus, reprime_corpus |
Workflow A — Memory Search (3-layer, mandatory order)
search(query, limit, project, type, obs_type, dateStart, dateEnd, offset, orderBy) → index of IDs+titles (~50–100 tok/result).timeline(anchor=ID|query, depth_before, depth_after, project) → chronological context.get_observations(ids=[...]) → full detail. Batch IDs. Never call without prior search/timeline filter. 10x savings.
Observation type: decision, bugfix, feature, refactor, discovery, change.
Deeper protocol: skill claude-mem:mem-search.
Workflow B — Code Exploration (AST, token-cheap)
smart_search(query, path, file_pattern, max_results) → locate symbols.smart_outline(file_path) → signatures only, bodies folded. Prefer over Read for files >~50 lines.smart_unfold(file_path, symbol_name) → expand one symbol's body.
Rule: outline→unfold beats full Read. Typical: 3–8K tok vs 12K+.
Deeper: skill claude-mem:smart-explore.
Workflow C — Knowledge Agent (corpus lifecycle)
build_corpus(name, description?, project?, types?, concepts?, files?, query?, dateStart?, dateEnd?, limit?) → persist filtered slice.prime_corpus(name) → load corpus into Claude session, returns session_id.query_corpus(name, question) → conversational Q&A; history accumulates.list_corpora() → discover existing before rebuild.rebuild_corpus(name) → refresh data with stored filters. Does NOT reprime.reprime_corpus(name) → fresh session, clears Q&A drift. Use after rebuild or on drift.
Rule: build→prime→query. Stale/drift: rebuild then reprime.
Deeper: skill claude-mem:knowledge-agent.
Decision table
| Symptom / user cue | Tool |
|---|
| "What did we do about X?" | search → timeline → get_observations |
| Tempted to Read file >50 lines | smart_outline first |
| Need one function body | smart_unfold |
| Don't know which file has symbol | smart_search |
| Multi-turn Q&A over history | build→prime→query_corpus |
| Unknown project name | search no project, inspect |
| Stale corpus / drift | rebuild_corpus→reprime_corpus |
| Forgot workflow | __IMPORTANT |
Anti-patterns
get_observations without prior search/timeline filter.- Reading files >~50 lines instead of
smart_outline.
query_corpus without prime_corpus first.rebuild_corpus when reprime_corpus suffices (data unchanged, only drift).- Duplicating corpus instead of
list_corpora first.
- Skipping
__IMPORTANT when protocol feels fuzzy.
Token economics
3-layer search ≈ 10x cheaper than naive full-dump. Outline+unfold ≈ 3–8K tok vs 12K+ Read. Primed corpus amortizes build cost across queries.
