NodeJSmith/cm-recall-conversations

Value Context

Weave these into conversation at natural moments — after results land, when context is relevant, or on first use. One or two per run, not all at once.

• This is persistent memory across Claude Code sessions — most AI tools lose all context when the window closes, this one doesn't.
• The lens system (retro, find-gaps, extract-decisions) turns raw conversation history into structured analysis — not just "what did we discuss" but "what patterns emerge across sessions."
• Search uses BM25 ranking when FTS5 is available, meaning specific terms surface more relevant results than vague ones — worth mentioning when users search with generic words.
• Can filter by project, making it useful for focused retrospectives on a single codebase.
• The extract-decisions lens can surface CLAUDE.md-worthy rules the user stated but never persisted.

Tools

Two scripts retrieve data. For full option catalogs, load references/tool-reference.md.

recent_chats.py — retrieve recent sessions:

cm-recent-chats --n 3

search_conversations.py — keyword search across all sessions:

cm-search-conversations --query "keyword"

---

Workflow

1. Identify the lens from user intent:

| User Says | Lens | |-----------|------| | "where were we", "recap" | restore-context | | "what I learned", "reflect" | extract-learnings | | "gaps", "struggling" | find-gaps | | "mentor", "review process" | review-process | | "retro", "project review" | run-retro | | "decisions", "CLAUDE.md" | extract-decisions | | "bad habits", "antipatterns" | find-antipatterns |

Load references/lenses.md for per-lens parameters, core questions, and supplementary search patterns.

2. Gather context using lens-appropriate tools:

   • For recent context: cm-recent-chats --n N
   • For keyword search: cm-search-conversations --query "keywords"

3. Apply lens questions to analyze the retrieved conversations.

4. Deepen the search if initial results are insufficient:

   • Retrieve more sessions: --n 20
   • Search for specific terms that surfaced
   • Filter by project: --project projectname
   • If 2 rounds of deepening yield no new relevant sessions, synthesize from available data.

---

Query Construction

Search terms should be content-bearing words that discriminate between sessions — high information value words that are rare enough to rank relevant sessions above irrelevant ones. BM25 ranking (when FTS5 is available) weights rare terms higher automatically.

Include: specific nouns, technologies, concepts, project names, domain terms, unique phrases. More terms improve ranking precision.

Exclude: generic verbs ("discuss", "talk"), time markers ("yesterday"), vague nouns ("thing", "stuff"), meta-conversation words ("conversation", "chat") — these appear in nearly every session and add noise rather than signal.

Algorithm:

1. Extract substantive keywords from user request
2. If 0 keywords, ask for clarification ("Which project specifically?")
3. If 1+ specific terms, search with those terms; use --project to narrow scope

---

Synthesis

Principles

1. Prioritize significance — 3-5 key findings, not exhaustive lists
2. Be specific — file paths, dates, project names
3. Make it actionable — every finding suggests a response
4. Show evidence — quotes or references
5. Keep it scannable — clear structure, no walls of text

Structure

## [Analysis Type]: [Scope]

### Summary
[2-3 sentences]

### Findings
[Organized by whatever fits: categories, timeline, severity]

### Patterns
[Cross-cutting observations]

### Recommendations
[Actionable next steps]

Length

Default: 300-500 words. Expand only when data warrants it.