Skip to main content
An agent that talks with someone over weeks needs to remember the conversation and recall the right pieces later. Infino stores each message once, its text alongside an embedding, and serves recall over that single table: by keyword, by meaning, by both fused together, and by SQL. Memory lives on a local path or object storage, so it grows with the agent’s history rather than a fixed cluster.

​
What memory needs from a store

A plain vector store gives you the first of these and nothing else. Keeping keyword search, semantic search, and SQL in one engine is what removes the separate keyword index, the rerank service, and the second database an agent memory usually needs.

​
Build it

The agent-memory example is a runnable implementation on the Node binding. It loads one real, months-long, multi-session conversation from LOCOMO, a public long-term conversational-memory dataset, turns each message into a memory timestamped with its session, and then demonstrates hybrid recall, SQL over memory, and deletion against it.
Start from the Quickstart for the shape of a table, embeddings, and search, then follow the example for the memory-specific parts.

​
Why Infino for agent memory

  • Hybrid recall in one engine. Keyword (BM25) and semantic (vector) matches are fused in a single query, so an agent finds memories by exact terms and by meaning at once, with no separate keyword index or rerank service.
  • Object-storage-native. Memory is stored as Apache Parquet on a local path or object storage; it scales with the agent’s history rather than a fixed cluster.
  • SQL over memory. The same store answers structured and time-based questions over an agent’s history.

​
See also

Last modified on August 31, 2026