Build update: what "memory you can audit" actually looks like.
The screenshot below is the part of the admin console I get asked about most.
Under the hood:
Every interaction is stored as an immutable episode. Nothing is ever overwritten.
Episodes are compiled into typed memories with confidence scores instead of raw text blobs.
Retrieval returns a ranked, token-bounded context bundle, so you control exactly what reaches the model.
Every retrieved fact links back to the episodes it came from.
That last step changes debugging completely. Instead of asking "Why did the agent say this?" and digging through logs, you can trace every retrieved fact back to its source in seconds.
Behind the scenes, every change runs through CI with 708 unit tests and 56 evaluation assertions to keep retrieval behavior stable.
Happy to go deeper on any part of the architecture. Anything you'd want to see before launch?
Replies