Turn blocked tasks, user corrections, fallback actions, failed tools, and human handoffs into a durable routing layer that helps agents recover better next time.
Exception memory only works when the record is structured enough to change future behavior. Store the recovery pattern, not just a transcript fragment.
The trigger says what broke. The attempted action says what the agent was trying to do. The fallback says what happened instead. The outcome says whether it worked. The reuse rule says what future agents should do differently.
Approval expired, tool failed, instruction was ambiguous, or human handoff was required.
Resolved, delayed, abandoned, corrected, escalated, or converted into a new policy.
Ask earlier, avoid a tool, require approval, retry later, or package context differently.
Add explicit events for approval timeout, user rejection, unclear instruction, failed browser action, broken asset, stale source, and human escalation. Do not infer every exception from vague failure text.
Store the channel, task, tool, approval state, and relevant link or artifact. Exception memory must be usable outside the original conversation.
Record what the agent did after the exception: saved a draft, retried later, downgraded scope, asked the user, or handed the task to a person.
Let users inspect, edit, and delete exception records. Recovery memory can include sensitive workflow context.
Use repeated exceptions to adjust prompts, route earlier to humans, change approval timing, or prefer a safer tool path.
"The exception ledger should answer one question: what should the next agent do differently?"Implementation note, personal agent operations desk
A detached memory table is not enough. The record should affect the next text approval, browser action, publishing run, or handoff.
Use text message AI assistant flows to capture corrections and approval exceptions.
Pair with computer-use cache for failed tools, auth gates, and risky actions.
In agent website workflows, remember stale deploys, failed links, and release gates.
Capture which context helped the person resolve the task fastest.
Should exception memory be automatic?
Capture can be automatic, but user review and deletion should be available.
What is the first exception to track?
Start with approval timeouts or failed handoffs because they usually create visible user friction.
Where does Supers fit?
Supers is relevant for personal AI agents that work across text, browser automation, publishing, and human escalation.
Exception memory turns recovery into reusable policy. Build the schema around the moments where the agent stopped, asked, retried, corrected, or handed work to a person.
Start with a small exception schema, connect it to the workflows that create failures, and use repeated patterns to update routing policy.