There isn't one memory, there are a few, each with a different job. This page maps them so it's clear what holds what, how long it lasts, and which one makes OpenLoop remember across months of use.
Think in three timescales. The model's attention holds the current moment, a run log holds one task end to end, and the durable stores hold everything worth carrying forward.
The working memory of a single session.
A replayable record of one task, start to finish.
Auto-captured episodes, recalled by meaning. The long-term brain.
Hand-curated facts about you and your projects.
The structure of a repo, queryable instead of re-read.
Who the agent is, plus its model of you. The Hermes three-tier pattern.
One thing often gets called memory but isn't: the secrets vault. That's configuration (keys and settings), not recall. It is deliberately kept separate.
The durable improvement comes from one loop running quietly under every task. Capture is cheap and never blocks the work; recall happens up front so the agent starts a task already knowing what it learned last time.
The confusion was having several stores with no clear division. The standard is simple: an always-on identity frame, three long-term recall layers that do not overlap, and the two short-lived ones above them.
Short-term context and the per-run log sit on top and feed the durable layers. Nothing else should call itself memory.