Procedural memory is now on by default
Remnic now captures how you do things, not just what is true, and injects the runbook when a matching task starts. No configuration required.
Facts are half of what an agent needs to remember. The other half is procedure: the order you deploy in, the checks you run before opening a PR, the way credentials get rotated in your setup. Agents relearn these from scratch every session, usually by getting them wrong once first.
Remnic’s procedural memory captures multi-step runbooks and injects them when a matching task begins. As of the current release line it is enabled by default.
How a procedure gets in
There are two paths. The direct one: you tell your agent the procedure, and extraction stores it as a category: procedure memory. The passive one: a trajectory miner watches for behavior you repeat across sessions, clusters it, and files a pending-review procedure candidate. Nothing self-approves. You review the candidate before it becomes active.
How a procedure comes out
At task initiation, Remnic matches the incoming task against stored procedures and injects a relevant-procedures block before the agent starts working. The agent begins the deploy already knowing your deploy order, instead of reconstructing it from the repo and hoping.
The safety posture
Default-on required raising the bar for what qualifies. Thresholds are conservative: weak matches inject nothing, and mined candidates wait in pending review until a human approves them. If you want procedural memory off entirely, set procedural.enabled: false in config.
The full behavior, including trajectory mining and threshold details, is in docs/procedural-memory.md. For how this plays with repo-scoped coding memory, see the coding agents page.