check-yo-self

Detect memory drift in agent transcripts by analyzing session evidence for repetition, missing receipts, and corrections. The evaluation compares recalled context against fresh evidence to identify when an agent relies on outdated information.

Transcript-based self-evaluation is an adversarial introspection process that scans session logs for repeated pattern copying and stale citations. It turns observable transcript evidence into a tiered verdict with a recommended next action.

Trigger a self-evaluation check by asking the assistant to run the analysis on the current session. It processes the transcript and returns a tiered recommendation along with supporting signals and next steps.

Run a drift detection check after long working stretches, repeated pattern-copying, or when a user signals fatigue and questions agent sharpness. It catches compounding errors before the session derails.

You can get a tiered recommendation that produces a green, yellow, orange, or red verdict. The evaluation advises whether to continue, slow down, refresh, or stop based on detected drift and correction frequency.

Self-evaluation works using only internal scripts with no external dependencies. It analyzes the existing session transcript directly to produce a freshness check and drift verdict without requiring additional packages.