agent-memory-systems

Use a memory framework like LangMem, MemGPT, or Mem0 to store semantic, episodic, and procedural memories in a vector database. Retrieve relevant memories at runtime by combining embedding similarity with metadata filters such as user_id and memory type.

Pinecone suits managed enterprise scale, Qdrant handles complex metadata filtering, Weaviate supports hybrid search, ChromaDB fits prototyping, and pgvector works if you already run PostgreSQL. Choose based on scale, filtering needs, and cost.

A general range is 256-512 tokens, but optimal size depends on content type: around 512 for documentation, 1000 for code, and 256 for conversations. Test multiple sizes against real queries and measure retrieval recall before committing.

Semantic similarity is not the same as relevance, so pure vector search can match topically similar but wrong-type content. Fix this by filtering on metadata like user_id and memory type first, using hybrid keyword search, and reranking with a cross-encoder.

Apply time-decay scoring so recent memories rank higher, update preferences in place instead of appending, and run conflict detection on storage using an LLM check. Periodic consolidation merges duplicate or conflicting memories into one current version.

No. Embedding models produce different vector spaces, so queries embedded with one model will not match documents embedded with another. Track the embedding model in metadata and re-embed all documents when migrating to a new model.