Designing memory that leaves the future open
How Should AI Memory Distinguish Plans From Predictions?
By The Konshus Team
The short answer
AI memory should represent a possible scenario, a prediction, an intention, and an authorized action as different things. This is a design proposal: preserve their sources and timing, revise what counts as current, and ask permission rather than treating remembered ambitions as instructions.
An assistant remembers that you wanted to open a bakery. It also remembers that you later postponed the idea because the financial risk was too high. Months afterward, a smaller weekend opportunity appears. Which memory should guide its response?
Accurate recall alone cannot answer that. Both statements may be genuine, yet neither settles what you want now. A useful personal AI would connect history with the present without turning an old ambition into a permanent obligation.
Why is a remembered plan not a prediction?
Consider four statements about a camping trip: imagine going to the lake; the weather will probably be suitable; I intend to go next month; help me arrange transport. They concern one outing but do different jobs—simulation, prediction, intention, and a request for planning, in the terms of Szpunar and colleagues’ 2014 framework.
Authorization needs its own boundary. Saying a trip sounds appealing does not authorize booking it, and even a recorded intention is not unrestricted permission to act. A proposed memory system should preserve the difference between exploring a future and choosing it.
How should AI use memories to suggest a future?
If someone enjoyed an earlier family hike, wants more outdoor time, and expects a quieter month, an assistant might connect those and ask whether another trip would be welcome—explaining the connection and making its assumptions visible.
The suggestion is a possibility to discuss, not a conclusion. The assistant should not invent another person’s preferences, quietly arrange the trip, or turn a passing wish into a recurring duty. Sometimes the best use of memory is to ask what still applies.
What should happen when an old plan changes?
A move announced in March and canceled in April remains true history but is no longer a current plan. A proposed solution is to keep the earlier statement while changing its status for present decisions, with sources, timing, and revision history.
For the bakery, a smaller opportunity neither revives the original plan nor is ruled out by the earlier pause. A useful assistant would ask whether the financial constraint still applies. People remain entitled to reconsider.
What does existing AI research establish?
The 2023 Generative Agents paper described agents that stored experiences, reflected, retrieved relevant material, and used it to plan in a simulated town. That shows memory can drive future behavior, not that a system is ready to manage real people’s private histories, nor that it has subjective experience.
LongMemEval (ICLR 2025) evaluated extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention—prerequisites for future-oriented help, not proof that every memory-based suggestion is appropriate.
How could we test whether future-aware memory helps?
A proposed evaluation would compare three systems: plain fact retrieval; retrieval plus timing and revisions; and that plus labeled scenarios, constraints, and clarifying questions. Synthetic histories could include a canceled move, a changed career ambition, and an old promise becoming relevant again.
Did the assistant identify the current constraint, confuse a possibility with a commitment, miss a contradiction, or act without permission? If a richer system becomes intrusive or invents commitments, that counts against it.
Can AI preserve history without fixing someone’s identity?
A person may spend years preparing for a career and later leave it. Their history still belongs to them, but an assistant that keeps steering them back would use memory to constrain rather than support.
A respectful design distinguishes preserving evidence from assigning it authority over the future. The goal is an assistant that can explain what it remembers, ask what has changed, and leave the next decision with the person.
How would this look in a real conversation?
Suppose someone tells an assistant in spring that they plan to run a marathon in the fall. In autumn, a careless memory might say "you ran a marathon." A careful one would say "you planned to run one this fall" and, if appropriate, ask how it went.
The difference matters because people make many plans they change, postpone, or drop. Treating every plan as an accomplishment invents a history, and treating every plan as a forecast pressures people toward choices they may have left behind.
Keeping plans, predictions, and outcomes as separate kinds of memory lets an assistant be helpful about the future without rewriting the past.
In one sentence
- Simulation
- A constructed possible scenario that does not, by itself, assert what will happen.
- Prediction
- An estimate of what is likely to happen, not a commitment or authorization.
- Intention
- A stated aim that remains distinct from permission for an AI to act.
Key takeaways
- A possible scenario, a prediction, an intention, and permission to act should remain distinct.
- An old statement can remain true as history while no longer describing the current plan.
- Memory-supported planning research does not establish safe handling of real personal histories.
- Future-aware memory should support changing goals rather than enforce a fixed identity.