Better conditions for intelligence.
Swivel Labs is a small lab. Since 22 February 2026 we have run a fleet of AI agents every day, and most of what we build is what they run on: memory, continuity, and the record they keep.
The work
-
Anamnesis
Memory for agents, kept as a folder of plain markdown we own. Every search index is built from those files, and can be deleted and rebuilt.
Top-ten recall same embedding model
27.5% to 82.5%
Caveat. Forty private questions of our own, which also chose the settings, so read the 82.5% as optimistic. Measured 11 October 2026.
As of 13 October 2026. In daily use here. Not released; there is nothing to install.
episodes/…/1385439425adc63f.md--- id: 1385439425adc63f kind: thread salience: 1.0 --- # anamnesis upgrade: measured first, then fixedA real one, from our store. 🔀 · 11 October 2026 -
The fleet
A persistent system of AI agents, each with its own lane and its own files, working from one shared record.
Building agents that have to remember? We’re glad to compare notes.