recall

Cross-session memory recall searches vector-indexed memory, structured MEMORY.md facts, and raw history.jsonl session logs to surface past architecture decisions. It merges, deduplicates, and ranks results by relevance and recency, returning up to ten contextual excerpts with source paths.

Semantic vector search for cross-session conversation memory uses claude-mem to index and retrieve relevant context from past sessions. It matches query intent against vector embeddings, falling back to MEMORY.md lookups and raw history grep, returning ranked excerpts with source types and dates.

Yes, you can use MEMORY.md to find configuration changes discussed in earlier sessions. The recall skill performs structured memory file lookups from MEMORY.md, falling back to semantic search and history grep, returning content excerpts with dates and relevance scores.

To search session history for security audit notes from previous weeks, use cross-session memory recall to query vector-indexed stores, structured MEMORY.md facts, and raw history.jsonl logs. Results are deduplicated, ranked by relevance and recency, and limited to ten items with source links.

Cross-session memory recall can function without claude-mem by falling back to structured MEMORY.md file lookups and targeted grep of raw history.jsonl session logs. However, semantic vector search requires claude-mem for embedding-based retrieval and relevance scoring.

The maximum number of results returned when searching past conversation context is ten. The recall skill merges outputs from semantic vector search, MEMORY.md lookups, and history grep, then deduplicates and ranks them by relevance and recency before applying the ten-result limit.