Similarity is not memory by itself
Vector search retrieves semantically similar text. That is useful, but it is not the same as remembering. A memory system still needs ingestion rules, metadata, freshness, conflict handling, and deletion or archival policies.
RAG for agents works best when retrieval is treated as an observation tool, not as a giant automatic context dump.
Retrieve with metadata
Store source, timestamp, author, confidence, and scope with each chunk. When retrieved memory conflicts, the model needs those fields to judge which memory should win.
Chunking matters too. Chunks should be small enough to retrieve precisely and large enough to preserve meaning. Blindly splitting every 500 tokens is a starting point, not a law.