The Behavioral Context Window gives Nomotic memory. Instead of evaluating each agent action in isolation, the context window looks at recent history to understand whether the current action is consistent with how the agent has been behaving.
Why it matters
An action that would be perfectly normal for an agent that has been doing similar things all week looks very different coming from an agent that has never taken that action type before. Without memory, governance can’t tell the difference. With the Behavioral Context Window, it can.
How it works
The context window looks at the last N evaluations for an agent (where N is configurable: 10, 25, 50, or 100) and computes:
Baseline rates — what percentage of recent actions were ALLOW, ESCALATE, and DENY, and what the average UCS score has been.
Trend direction — whether the agent’s UCS score and allow rate are improving, declining, or stable over the window period.
Anomaly score — a 0 to 1 measure of how different the agent’s recent behavior is from its own established baseline. Scores above 0.3 warrant attention. Scores above 0.4 indicate a significant behavioral shift.
Dominant action types — the top 3 action types by frequency in the window.
Pattern flags — human-readable observations such as high escalation rate, low confidence scores, or behavioral shift detected.
Configuring the window
You can configure the window size per agent from the agent detail page in the Governance tab. The options are 10, 25, 50, or 100 evaluations. A larger window gives a more stable baseline but is slower to reflect recent changes. A smaller window is more responsive but more sensitive to short-term variation.
For most agents, 50 is the right default. For high-volume agents with stable behavior, 100 gives a more reliable baseline. For new agents or agents in active development, 25 gives faster feedback.
Viewing context
The Behavioral Context Window for each agent appears on the agent detail page under the Governance tab. Fleet-wide context for all agents is visible in Intelligence → Fleet Context.