Adding to Shaked’s point: persistent memory can become stale and deciding what to retain and trust is a security trade-off. The current stateless approach, combined with the code sandbox, can rebuild context quickly. The idea is to add smarter memory without losing that property.
Comments
Thanks! in the works, you can partly go around this now by piping out and back in but we'd like to make it smarter - llm wiki and graph.
Adding to Shaked’s point: persistent memory can become stale and deciding what to retain and trust is a security trade-off. The current stateless approach, combined with the code sandbox, can rebuild context quickly. The idea is to add smarter memory without losing that property.