Preference memory helps an agent personalize. Exception memory helps it recover: blocked tasks, rejected actions, corrected assumptions, timeout fallbacks, and human handoff outcomes.
Most memory systems emphasize preferences: tone, favorite tools, personal facts, recurring workflows. Exception memory stores the operational evidence that teaches an agent when a path failed and what should happen next time.
It captures the action that was attempted, why it stalled or failed, what the user corrected, which fallback ran, and whether a human handoff resolved the task. That record prevents the agent from repeating brittle behavior under a new prompt.
Remember where approval latency killed usefulness and which fallback worked.
Store the correction as a policy update, not a transient chat message.
Track when a person fixed the task and what context made the handoff faster.
Record the exact condition that made normal automation unsafe or unhelpful: missing permission, stale context, tool failure, ambiguous instruction, policy boundary, or user rejection.
Store what the agent was trying to do in plain operational language. A future agent cannot learn from a vague failure label.
Capture what happened instead: draft saved, scope reduced, user clarified, approval timed out, task handed off, or workflow stopped.
Mark whether the fallback solved the task, delayed it, created extra work, or required a new policy.
Convert the exception into a future routing instruction: ask earlier, avoid this tool, require human approval, retry later, or package context differently.
"The failures are the training set. Exception memory turns them into routing policy."Research note, personal agent operations desk
Exception memory matters most when agents move through channels where context is easy to lose: text, browser automation, publishing, and human handoff.
Use text message AI assistant history to remember skipped approvals, unclear asks, and resolved corrections.
Pair computer-use cache with exceptions for failed selectors, login gates, and risky actions.
In agent website workflows, remember failed assets, broken links, stale deploys, and approval misses.
Remember what context the person needed so the next handoff starts closer to resolution.
Is exception memory the same as long-term memory?
No. It is a narrower operational ledger for abnormal outcomes, corrections, and handoffs.
Should every failure be remembered forever?
No. Store reusable patterns, keep audit context, and allow user-controlled deletion.
Where does Supers fit?
Supers is relevant for personal AI agents spanning messaging, browser work, publishing, and human escalation.
Personal AI agents become more reliable when they remember exceptions as reusable routing evidence. That is different from remembering preferences; it is remembering how work recovered.
Start by logging the exception fields that change future routing decisions. Then connect them to text, browser, publishing, and handoff workflows.