Memory tells an agent what happened. It doesn’t tell it what’s allowed to change. The standard is the second half of that sentence — this page is the first half’s home.
What exists
A working agent-memory architecture, running every day: memory as files you own, state that stays current, reasoning that appends and never silently rewrites, an index every session reads first. Not a proposal — the system that keeps this whole operation coherent across sessions, machines, and models.
- The map: the memory-flow map — the working architecture as one explorable screen, content-addressed on the shelf of maps.
Where it goes
Durable, governed agent memory: memory with weights, where what’s remembered carries consequence, what’s foundational can’t be silently rewritten, and what gets forgotten is a decision — not an accident of a context window. Governance and memory converge into one layer, because the question they answer is the same question.
A full thesis on this layer is queued — built from the running system, not from theory. It joins the map of the series when it’s real.
Direction, not dates. The parts ship as they’re proven, and their names arrive with them.
Govern your truth; keep your memory. Two halves of one answer.