The memory system has been redesigned to use LLM-based retrieval instead of grep-based keyword search. Previously, memories were found by extracting search terms and grep-matching against markdown files. Now:
Memories are listed with filenames and descriptions
An LLM selects the most relevant memories (up to 5) based on the user's query
Memories already read via the Read tool in the current conversation are filtered out to avoid duplication
Memory files support a description frontmatter field for better retrieval
Evidence
Selection prompt "You are selecting relevant memories for a user's query..." and function x8q for deduplication (search for "You are selecting relevant memories")
Strings lifted out of the shipped bundle, so the claim above can be checked against them.
