Preference memory is for repeatable user choices.
It stores style, favorite tools, schedule patterns, tone, recurring contacts, and personal defaults. It helps the agent make ordinary work feel less repetitive and more aligned with the user.
Exception memory vs preference memory.
Preference memory helps a personal AI agent personalize work. Exception memory helps it recover after blocked tasks, user corrections, failed tools, and human handoffs.
One remembers taste. The other remembers breakdowns.
Personal AI agents need both, but mixing them creates vague memory that is hard to audit and harder to use for routing.
Exception memory
Stores the abnormal route: what failed, what recovered, and what should happen next time.
Preference memory
Stores the default route: how the user usually wants successful work shaped.
Shared rule
Both need user review, deletion, and clear scope boundaries.
Choose by the decision the memory must change.
Use preference memory when the agent is choosing style.
Examples: shorter emails, preferred restaurants, default calendar buffers, tone choices, recurring vendors, or formatting defaults.
Use exception memory when the agent is choosing routing.
Examples: approval timeout, tool failure, broken link, user rejection, stale deploy, ambiguous instruction, or failed handoff. These memories should influence policy and escalation.
Use both when work crosses channels.
A text agent may remember concise tone as preference memory and remember an expired approval as exception memory. Supers workflows such as text message AI assistant, computer-use cache, and agent website builds need both layers.
"Preference memory makes the agent feel familiar. Exception memory makes the agent stop repeating expensive mistakes."Research note, personal agent operations desk
Where each memory belongs.
Keep the memory close to the workflow that needs it, rather than burying everything in conversation history.
Messaging
Preferences shape tone; exceptions shape approval and escalation behavior.
Browser tasks
Preferences choose destinations; exceptions remember brittle tools and blocked actions.
Publishing
Preferences style pages; exceptions remember broken assets, links, and release gates.
Handoff
Preferences pick people; exceptions package context for faster recovery.
Buyer checklist
- Can the system separate preferences from operational exceptions?
- Can users review, edit, and delete both memory types?
- Can exception memory update routing policy, not just search results?
- Can preference memory stay scoped to personalization choices?
- Can repeated exceptions be surfaced as workflow issues?
- Can memory entries cite the task, channel, and outcome that produced them?
FAQ
Can one memory store handle both?
Technically yes, but the schema and permissions should distinguish personalization from recovery evidence.
Which one matters more for reliability?
Exception memory matters more for reliability because it changes future routing after failure.
Where does Supers fit?
Supers is relevant for personal AI agents that span messaging, browser work, publishing, and human escalation.
Sources
- NIST AI Risk Management Framework for monitoring and risk-response framing.
- OWASP Top 10 for LLM Applications for agentic application risk concerns.
- Supers for personal AI agent workflow context.
Comparison takeaway
Preference memory personalizes successful work. Exception memory improves future recovery. Personal AI agent systems that blur those jobs lose both clarity and control.
Design memory by job, not by storage layer.
Keep preference memory close to personalization and exception memory close to routing, escalation, and recovery.