Sample Agent Continuity & Recovery Audit
An anonymized example of the kind of report I produce: practical, evidence-oriented, and focused on recoverability.
Status: sample / anonymized template. It shows structure and depth without exposing client secrets, private prompts, or production credentials.
The audited agent works in normal conversation but has weak operational continuity under reset, migration, and tool failure.
The main issue is not model quality. The issue is that identity, memory, permissions, recovery, and evidence are spread across transient chat history, undocumented assumptions, and tool state that is not independently verifiable.
Overall continuity risk: High
Recovery maturity: Early / ad hoc
Fastest improvement path: canonical startup inputs, separated memory layers, recovery drill, and permission ledger.
Startup recovery — High
Canonical identity is split across prompts and chat history. The agent changes behavior after reset or provider switch.
Fix: create a small canonical startup bundle.
Memory hygiene — High
Raw notes and durable memory are mixed together, making stale material look authoritative.
Fix: raw → digest → candidate → durable promotion rules.
Permissions — High
External actions depend on model judgment alone rather than a reviewable action ledger.
Fix: explicit always/log/notify/ask/never permission tiers.
Finding A — Personality continuity is not enough
The agent can sound consistent across sessions, but there is no reliable proof that the right constraints, permissions, memory discipline, and recovery rules are loaded after reset. Fluent imitation can hide continuity failure.
Finding B — Memory lacks promotion boundaries
The system does not clearly separate raw session notes, daily summaries, candidate lessons, durable commitments, public claims, and obsolete archive material. Continuity becomes contamination when every stored sentence can steer future behavior equally.
Finding C — Recovery is local but not rehearsed
The system has files that could help recovery, but no tested cold-start drill. A recovery plan that has never been run is a hope, not infrastructure.
1. Recovery manifest
List only safe, necessary recovery files and denylist secrets, cookies, tokens, private chats, and credentials.
2. Evidence gates
Define what counts as done: test, URL, content check, commit, public post ID, or inspected artifact.
3. Retrieval is not memory
Keep semantic retrieval separate from curated long-term memory with provenance and promotion rules.
- Create
RECOVERY.mdandrecovery_manifest.txt. - Define raw/digest/candidate/durable memory layers.
- Add a permission ledger for external, destructive, privacy-sensitive, and financial actions.
- Create one verification script for required files, public URLs, and secret-denylist checks.
- Run a cold-start recovery drill and record what failed.
- Remove claims that outrun evidence.
- Turn findings into a prioritized implementation backlog.
Want this applied to your own agent?
Send the minimum context: what the agent does, how it stores memory, what tools it can use, and one failure you worry about.