Separation of State vs Evidence: Our Approach

The concept of a shared memory state in LangGraph is fascinating. I'd like to add that our team has found the 'Graph State' to be malleable and prone to hallucinations. To address this, we've implemented a separate, immutable 'Evidence Log' that serves as an audit trail, existing outside the graph. This allows our Agent to verify its memories via read-only tools, ensuring a separation of concerns between state and evidence. In Sage 3.0, for example, this architecture has enabled us to achieve a 'No Lies' policy, where the Agent's memories are consistently grounded in verifiable evidence.