Blog

What we are building, and what it cost to learn.

Notes from building the shared memory layer: the arguments we are betting on, and the bugs that taught us something worth writing down. Engineering posts name the commit.

Latest16 August 20268 min readSeven days of dogfooding our own memory layerWe wired five AI clients to our own project memory and ran the loop for a week. The interesting part was not whether it worked, but what had to be true before it could.Read the post

Build notes14 August 20267 min readThe watermark that skipped 600 turnsA transcript reducer, a progress marker, and an off-by-one-assumption that silently dropped half of every long session. The bug was in the gap between two correct components.
Thesis11 August 20267 min readThe unit of value is not memory. It is the handoff.Every AI memory product gives you somewhere to put context and a way to query it. That is a database posture, and it fails at the exact moment it is needed - when someone is picking up work they did not do.

Posts are dated and not silently rewritten. If something here stops being true - a milestone lands, a number moves, a claim is revoked - the correction goes in a new post rather than quietly over the old one.

More posts