Tamper-evident AI memory
Show what your agent knew, and when it knew it.
An AI agent that stores memory can be asked, later, what it believed at the moment it acted. Ordinary memory cannot answer that question honestly — a row can be edited, and nothing about the row proves it wasn't. CapchaBrain seals every memory write into the same append-only chain that carries the rest of your evidence, so the answer is checkable rather than asserted.
Live at brain.capchacloud.com. Memories are written through /api/v1/brain/memories and retrieved or searched through /api/v1/brain/search.
What gets sealed
Each memory is sealed as a canonical record of the claim at seal time: the tenant it belongs to, the memory's identifier, the claim itself, its value, and the source it came from. The sealed artifact is content-addressed, so the same claim always produces the same digest, and any later edit produces a different one.
The seal is appended to the hardened audit chain as a brain_memory_sealed event. That chain is the same one published at /api/v1/transparency — its tip only ever extends, and anyone can record the tip over time and confirm it never rewrites.
The honest part: sealing is asynchronous
A write to CapchaBrain returns immediately with anchored: false. The seal is queued and the receipt arrives moments later. We report it this way on purpose: a write that claimed to be anchored the instant it returned would be claiming something that had not happened yet, and this is a product about not doing that.
What that means in practice: CapchaBrain is designed for showing what an agent knew after the fact, not for gating a decision on the seal in the same request. If you need a synchronous decision backed by evidence, that is what CapchaShield and the compliance envelope are for.
Tenancy and refusal
A memory is only sealed under a tenant that actually exists. The seal path verifies the tenant before writing anything into the permanent vault, so a bug or a compromised caller upstream cannot plant a record under an identity that was never provisioned. Unauthenticated calls are refused rather than silently dropped.
Where this is today
CapchaBrain is deployed and serving. It has not been through a third-party security review, and there are no production customers using it yet — the same disclosure that applies to every part of CapchaCloud, stated on the products page and in the Trust Center rather than in a footnote.