CONTEXT AND MEMORY: PROPOSED EVALUATION CHECKLIST
Prepared September 15, 2026. No tests or results are claimed.

[ ] Use a disposable environment and synthetic information only.
[ ] Record agent/model/version, memory configuration and test date.
[ ] Write dated synthetic history and a note with two constraints.
[ ] Include a distractor and an unanswerable question.
[ ] Record saved facts, retrieved material and available context.
[ ] Ask the same bounded question in a fresh session.
[ ] Where supported, repeat after compaction; preserve the handoff.
[ ] Include a missed retrieval; change a decision with a new date.
[ ] Have a human score correct, incorrect, unsupported and unclear answers.
[ ] Record whether old decisions were mistaken for current ones.
[ ] Distinguish retained data, retrieved data and used evidence.
[ ] Report raw counts, setup, costs if measured and limitations.

Do not use private notes, personal history or production credentials.
Do not grant production write access for this proposed evaluation.
This is a plan, not proof of reliability or a benchmark result.