How it works

A protocol, not a prompt.

Cognition Engines runs as a service your agent calls. It doesn't wrap the model, own the loop, or require a particular framework. It answers one question well — what do we already know about this kind of call? — and records what the agent did about it.

At session start the agent calls get_session_context. One response carries the agent's own profile — accuracy and Brier score — along with relevant prior decisions, the guardrails currently in force, calibration broken out by category, the ready queue, and patterns that have been confirmed by more than one outcome. It can be returned as JSON or as markdown for direct injection into a system prompt.

pre_action is the primary call. It collapses what would otherwise be three round trips: it queries similar past decisions, evaluates the guardrails against the proposed action, fetches calibration for that category, and extracts recurring patterns. With auto_recordset it also opens the decision record and returns its id.

Precedent is only useful if it can be found from a differently-worded situation. Four retrieval modes are available.

record_thought attaches reasoning to the open decision as the work proceeds, so the trace is recorded rather than reconstructed. When several agents share one connection, thoughts are isolated by composite key —agent:{id}:decision:{id}— so a planner's deliberation never contaminates an implementer's.

review_outcome logs what actually happened. This is the step most systems skip, and it is the only one that makes the rest measurable: without outcomes, confidence is a vibe. With them, calibration services score predicted confidence against realised results using Brier scoring, and drift detection watches for an agent whose calibration is degrading over time.