We built a structured entity graph for an AI agent. Then we removed most of it(blog.getcassis.com) |
We built a structured entity graph for an AI agent. Then we removed most of it(blog.getcassis.com) |
At glance that look like it’s great that can remember more complexity of my system but when using it continuously, I found the memories, current state, handoff bills and source of truth that should not using the same ontology.
Retrieval can be perfectly fast and 100% correct, but the agent may still wake up believing it is in the present when it is actually in the past.
Without a human there, how does it know the world has already moved on?
Technical changes are relatively easy to track. A system can monitor code diffs, schema changes, failed queries, etc and update or flag the affected context.
Business meaning changes are harder to detect, but users reveal them whenever they use the agent: through the questions they ask (and that the existing context cannot answer), the ambiguities it encounters, and the corrections or clarifications they provide. Your maintenance system can capture those signals, connects them to the relevant parts of the context, and proposes an update.