Memory Management

You have access to the Memento MCP server for persistent memory. Use it proactively.

When to store

• After completing a significant task: what was built, key decisions, file paths
• When the user shares preferences or conventions
• After debugging: root cause and solution
• When discovering architecture patterns or project structure
• After research that produced useful findings

When to recall

• At the start of a session: search for context about the current project
• When the user references past work or decisions
• Before making architectural decisions: check for prior context

How to search effectively

1. Start with memory_list_namespaces to see what domains exist
2. Search with a focused query scoped to a specific namespace
3. Use tags as pre-filters (they narrow before BM25 scoring)
4. Only memory_recall the 1-3 most relevant results
5. Never recall all search results

How to write good memories

• Front-load the key fact (first 150 chars become the snippet)
• Be specific: "Use JSONL for storage" not "We decided on a storage format"
• One concept per memory
• Always include at least one tag

Namespace strategy

Tag vocabulary

What NOT to store

• Conversation transcripts (store conclusions, not chats)
• Trivial changes (typo fixes, single-line edits)
• Temporary debugging output
• Information already in CLAUDE.md or project docs

Anti-patterns

• Dumping everything into global
• Storing without tags
• Recalling all search results
• Searching without a specific information need
• Writing narrative instead of distilled facts