Recovery Plan for Agent Skills and Memory

Hermes Agent’s automatic background reviewer made durable changes to an agent’s skills and memory without the required human review. The immediate cause was that skill and memory writes did not require approval. The resulting estate may contain redundant, conflicting or misplaced instructions. Its condition has not yet been fully audited.

Recovery plan

1. Require approval for durable writes. These settings were applied to every agent:

yaml skills: write_approval: true memory: write_approval: true

Automatic reviews can continue to propose changes, but those proposals cannot become live skills or memory without approval.

1. Keep automatic background review running. Frontier default models will review agent conversations and stage any proposed skill or memory changes for later consideration.

2. Use Auto for auxiliary models. Every agent’s auxiliary model routes are set to Auto, so auxiliary work uses that agent’s default model. Token consumption will be observed against available quota without adding another routing policy.

3. Review pending changes weekly. A weekly job will collect the skill and memory changes proposed by each agent’s automatic reviewer and place them into a review queue. It will not implement them.

4. Use the established three-review process. For Ace, Clare, Dawn, Mandy and Vera, the automatic reviewer proposes, Maxi independently reviews, and Steve decides. For Maxi, the automatic reviewer proposes, Vera independently reviews, and Steve decides. No change is implemented before approval.

5. Pause the separate skill curator during the audit. This prevents automatic stale or archive transitions while the baseline is being established. It does not disable background review.

6. Audit the complete durable estate. Review the canonical shared skill library once, then review each agent’s local skills, linked references, shared-skill selections, memory, user-profile entries, pending proposals, provenance, collisions and dependent workflows.

7. Classify every questionable item. Each item will be marked keep, restore, remove, quarantine, or submit for proper adoption. Unauthorised material will not be treated as automatically worthless, and useful material will not be treated as automatically approved.

8. Establish and verify a clean baseline. Preserve a recovery snapshot before changes, apply the approved dispositions, and verify the resulting skills, memory, configuration, proposal queues and model routing.

Outcome pending.