recall
Semantic memory recall using FTS5 and Chroma. Use PROACTIVELY whenever the user asks about past events, themes, conversations, or when historical context would enrich a response. Trigger words: remember, when did we, what did we talk about, last time, before, previously, that conversation, that time, history, past, recall.
Installation
Follow the repository instructions for this skill and your agent. Review the files and dependencies before adding it to your workflow.
View source ↗GuidesSkill instructions
Semantic Memory Recall
This skill searches through semantic memory indexes for associative recall of past conversations, themes, and context.
When to Use This Skill
Use proactively when:
- User asks about past conversations or events
- User references "that time when..." or "remember when..."
- Historical context would enrich your response
- User asks about themes discussed previously
- You need to recall what was said about a topic
Two Memory Systems
1. SQLite FTS5 (Keyword Search)
Fast keyword matching with BM25 ranking.
~/.claude-mind/system/bin/memory-index search "query" [limit]
Best for: Specific terms, names, dates, exact phrases
2. Chroma Vector Database (Semantic Search)
Embedding-based similarity search - finds related content even with different wording.
~/.claude-mind/system/bin/chroma-query "query" [n_results]
Best for: Themes, concepts, "conversations like X", finding related discussions
Search Strategy
-
Determine query type:
- Specific term/name/date → Use FTS5 first
- Theme/concept/vague reference → Use Chroma first
- Comprehensive search → Use both
-
Run searches:
# Keyword search (FTS5)
~/.claude-mind/system/bin/memory-index search "coffee shops" 10
# Semantic search (Chroma)
~/.claude-mind/system/bin/chroma-query "conversations about morning routines" 5
# For comprehensive recall, run both
~/.claude-mind/system/bin/memory-index search "project planning" 5
~/.claude-mind/system/bin/chroma-query "discussions about work priorities and goals" 5
- Synthesize results: Combine findings from both systems, noting dates and context.
Output Format
When presenting recalled memories:
- Lead with the most relevant finding
- Include dates to establish timeline
- Quote key excerpts when helpful
- Note which system found each result (keyword vs semantic match)
- Connect findings to the user's current question
Examples
User: "What did we talk about when I was at that coffee shop?"
# Semantic search for coffee shop conversations
~/.claude-mind/system/bin/chroma-query "coffee shop conversation" 5
# Keyword backup
~/.claude-mind/system/bin/memory-index search "coffee" 5
User: "Remember that discussion about Q1 planning?"
# Semantic for the theme
~/.claude-mind/system/bin/chroma-query "Q1 planning discussion goals priorities" 5
# Keyword for specifics
~/.claude-mind/system/bin/memory-index search "Q1" 5
User: "What have I said about my work situation?"
# Broad semantic search
~/.claude-mind/system/bin/chroma-query "work job career situation feelings" 10
# Follow up with specific terms found
~/.claude-mind/system/bin/memory-index search "specific_term_found" 5
Data Sources
Both systems index:
- Episode logs (daily conversation records)
- Reflections (dream cycle outputs)
- Learnings
- Observations
- Decisions
- Person profiles
Results include source type and date for context.
Examples
Read the usage examples in the source repository.
View source ↗Source history
- Indexed on
- 2026-01-20
- Index updated
- 2026-01-28
- Branch
- main
Index and sync dates describe this directory. View the source repository for code history.
GitHub ↗chat Comments (0)
Sign in to join the discussion and leave a comment.
No comments yet. Be the first to share your thoughts!