VERIFYAgents & automation
Invalidate: Agent Memory Management
Tejas Chopra@chopra_tejas
Invalidate is a system that manages agent memories by checking the validity of stored memories against new events, marking stale memories without rewriting them.
Jumping on the bandwagon of building something with @typesafeai #jev model
Agents remember. They never un-remember.
We built invalidate: every time something happens, every stored memory is asked one question. Is this still true? Stale ones get marked. Nothing is rewritten.
Why nobody did this: it is memories × events checks.
With an LLM that is too slow and too expensive, so memory products check the ten nearest, or nothing.
With Jev a check is 150 ms and $0.00006. So we check all of them.
That is how we think about Jev. Not a smaller LLM. A model that answers a yes/no question with a probability, fast enough and cheap enough that code can ask it thousands of times and own the decision itself.
Open source, works in front of @mem0ai , @trychroma , @LangChain , or a Markdown file.
github.com/chopratejas/invalidate