AI agents
Persistent memory and tool retrieval for agents that run longer than one prompt.
The problem
Agent frameworks forget what they did yesterday, or they stuff the entire transcript back into the model.
How retrieval closes it
Store memories as points with a type, a user, and a timestamp. Retrieve the few that match the current task, not the whole log.
Three features that matter
- Payload filters for user, agent, and memory type
- Real-time writes so the next turn sees the last tool result
- Multivector memories: summary plus raw text