Can be used for managing long-term memory in AI agents?

All answers to this question.

is perfect for agentic memory because it supports an upsert operation. This allows you to either insert a new memory or update an existing one based on a unique ID. For long-term memory, you should structure your metadata with timestamps and "importance" scores. This way, your agent can query for the most relevant and most important memories. The indexing is incremental, meaning you don't have to re-index everything; the HNSW graph updates as you add data, keeping the agent's "thinking" process fast and responsive even after thousands of interactions.

   Answered 2025-01-07 by Sandra Robinson

Is there a way to weight more recent memories more heavily during a similarity search, or is it purely based on the vector distance?

   Answered 2025-01-09 by Gary Walker

• Gary, the distance metric is purely mathematical, but you can achieve "recency weighting" through two-stage retrieval. First, fetch a larger set of results from the database, then apply a decay function to the scores based on the timestamp in your metadata before presenting the final context to your AI agent.

     Commented 2025-01-10 by Donald Young

I use it to store user-specific profiles, and the metadata filtering makes it easy to isolate memories for different users.

   Answered 2025-01-12 by Karen Allen

• That isolation is key, Karen. By filtering by user_id in the metadata, you ensure the agent never accidentally retrieves a memory belonging to a different person.

     Commented 2025-01-13 by Sandra Robinson