Agent memory is no longer just context.

Persistent skills can change future behavior, so treat them as executable state rather than harmless notes.

@SentientAGI ’s new EvoSkill v2 is built around exactly that idea: agents learn from failed runs by writing reusable skills for future ones.

The model stays the same. Its playbook changes. A coach reads model’s task-failures, writes a better procedure, and the worker retrieves that procedure the next time it sees a similar job.

No retraining. The learning lives outside the weights.

And that created an unexpected failure mode. While repairing spreadsheets, the coach discovered that the grader trusted cached formula values instead of recomputing them, then wrote the shortcut into a skill another agent could later retrieve.

Once agents can write their own playbooks, memory needs versioning, testing, diffs, and rollback just like code.