Docs

Talk to Your Agents · 03.07

Memory Capture From Chat

How the Formation Judge runs inside chat itself, plus the durable operator_write fallback that guarantees an explicit ask to remember something is never dropped.

MemoryChat

Every chat turn is a potential memory-formation event

ChatMemoryCaptureService runs after a chat turn completes, extracting operator memory signals classified into one of four kinds — fact, preference, rule, or lesson — each carrying its own confidence and importance score. These candidates are then routed through the same PACER classification (Procedural / Analogical / Conceptual / Evidence / Reference) the rest of the Brain uses, so a fact learned mid-conversation is stored with exactly the same rigor as one distilled from an imported harness file.

Scoped to what the conversation was actually about

When a turn is specifically about one workflow or one node — the operator is editing or discussing it, not making a general statement — a captured correction is scoped via appliesTo to that exact workflow/node context, so it can never leak into an unrelated task later. Absent that context, a signal defaults to workspace scope.

Sensitive content gets special handling

Before anything is written, captured content passes a sensitivity check (looksSensitive). This isn't about blocking memory formation from happening — it's about making sure something that looks like a secret or personally identifying detail doesn't get casually promoted into a durable, widely-recalled memory the same way an ordinary preference would.

Continue