Most of what we build runs on an AI agent harness — the rules, skills, and memory that let the assistant pick up a project without being re-briefed every time. The single piece of it we're proudest of isn't how it remembers. It's how it forgets. We extracted that piece, cleaned it of anything Dekimu-specific, and published it on GitHub under an MIT licence. It's our first public repo, and it's the start of putting a few of our tools out in the open.
The naive way to give an assistant memory is an append-only pile of notes. The problem shows up a month later: a fact you confirmed once sits there at full confidence forever, right next to a guess you jotted down and never checked. The pile fills with half-true, never-revisited claims, and you quietly stop trusting any of it. The usual fix — expire notes after N days — treats a core rule and an offhand observation as if they go stale on the same morning. They don't.
Our version gives every note a confidence band a human sets by hand, and a decay curve that erodes that confidence over time since it was last confirmed — faster for fragile project state, slower for durable working preferences, and graded so a strong fact survives far longer than a tentative one of the same age. The twist: the decayed value is never written to disk and nothing is ever auto-deleted. Decay only surfaces a note for review. A human then decides — confirm it (which resets the clock), weaken it, or retire it. The machine flags; the person judges.
The hard part of memory isn't storing things. It's noticing, on your behalf, which of the things you stored have quietly stopped being true.
It's a few hundred lines of zero-dependency Node, a drop-in skill, and a session hook — plus a design write-up that's honest about what we deliberately left out (no database, no embeddings, no continuous scoring). It lives at github.com/dekimuhq/claude-memory. This is the first of a handful of internal tools we plan to open up: the ones general enough to be useful to anyone, and self-contained enough to stand on their own. If it saves someone else's agent from drowning in its own stale notes, that's the point.