Auto-consolidate similar memories into generalized principles

[toejough]

Problem

Many memories are narrow observations of the same underlying principle. For example, 9 separate memories might each note a specific instance where code should have been deduplicated or reused. Individually they're noisy — each surfaces in slightly wrong contexts. But the extracted principle ("prefer reuse over duplication") would be high-value and broadly applicable.

This is a common pattern: multiple low-generalizability memories that share a theme but were created independently across sessions. They clutter surfacing, dilute each other's signal, and none individually captures the real lesson.

Desired behavior

Periodically (e.g., during maintain or a new consolidate subcommand):

1. Cluster memories by semantic similarity (embeddings or keyword overlap)
2. Identify clusters where N+ memories (e.g., 3+) share a common theme
3. Propose a single generalized memory that captures the principle
4. On approval: create the generalized memory, archive or remove the originals
5. The generalized memory should have higher generalizability score and broader (but still precise) keywords

Examples

• 9 memories about specific deduplication opportunities → 1 memory: "extract shared logic when 3+ call sites exist"
• 5 memories about forgetting to check git status before destructive ops → 1 memory: "always check VCS state before destructive operations"
• 4 memories about specific test isolation failures → 1 memory: "parallel tests must not share mutable state"

Design considerations

• Clustering could use TF-IDF similarity (already in the codebase) on memory content
• Haiku can propose the generalized principle from a cluster
• Should be dry-run capable (show proposed consolidations without applying)
• Threshold for cluster size should be configurable (default 3)
• Generalized memory inherits the combined feedback history of its sources
• Original memories get archived (not deleted) so consolidation can be reviewed/reversed

Related

• Retroactive generalizability scoring for existing memories #360 (retroactive generalizability scoring) — complementary; scoring identifies low-generalizability memories, this consolidates them
• Feedback-driven keyword refinement #346 (feedback-driven keyword refinement) — complementary; consolidation produces better keywords naturally
• Skill tier 2 promotion in CLAUDE.md already describes this pattern ("Generated from memory clusters, size >= 3")