News, announcements, and technical deep dives from the Aethelas team.
We're proud to announce the launch of Aethelas — a trust layer for verifiable AI memory. Learn, remember, prove, and unlearn with cryptographic guarantees. The platform is live with 20 active nodes, 84/84 passing tests, and 18/18 red-team attacks blocked.
Our latest technical report on provable data deletion. We demonstrate selective unlearning with cryptographic proof of erasure — removing one contributor's data changed the root hash and left other data bit-for-bit intact.
Read MoreWe test the harder, more realistic setting of class-incremental learning and report our findings honestly — including the limitations. All three replay conditions collapsed to near-chance accuracy.
Read MoreWhy we believe verifiable memory is the foundation for trustworthy AI systems in regulated industries. A look at the principles behind Aethelas and why transparency matters.
Read MoreWe've expanded the Aethelas website with 4 new pages: Security, Roadmap, Careers, and Feedback. Explore our security practices, see our future plans, join our team, and share your thoughts.
Read MoreA 561-parameter model beat a 350× larger one by 8.6 points when retrieval was perfect — but the advantage inverted past 17% retrieval error. The transferable result: measure retrieval accuracy first.
Read MoreAutonomous self-improvement is real but bounded by whether the task exposes a gradient. We measured the loop's behaviour across three deliberately different task types.
Read More