Updating the present without deleting the past
How should AI handle changing facts?
By The Konshus Team
The short answer
AI should distinguish corrections from changes over time, differences in context, competing perspectives, and unresolved conflicts. A proposed evidence-aware memory design would update the current view while preserving sources and relevant history, rather than assuming the newest statement always wins.
Imagine telling a personal AI that you are applying for a job in another city. Later you report an offer, then decline it, then reconsider moving under different conditions. A summary that simply says you want to relocate misses the sequence. A summary that says you never wanted to move is worse.
The problem is not a shortage of stored statements. It is a failure to represent change. A useful assistant should answer what your plan is now and explain what you planned before, without treating a once-genuine intention as an error. The designs below are proposals to test.
What can memory science contribute to AI design?
Memory research suggests a useful question: does new evidence revise an existing understanding, or does it belong to another situation? In a 2024 rat study, Kennedy and colleagues interpreted their results as supporting an account in which smaller prediction errors favor updating an existing memory state. Gershman and colleagues’ 2017 computational account frames learners as inferring hidden causes behind observations.
Neither tells engineers to copy a biological mechanism. The useful connection is organizational: contradictory information may indicate a correction, a changed situation, or missing context. Software needs ways to tell those apart.
What kinds of contradictions should AI distinguish?
A practical design proposal separates five kinds of disagreement—engineering categories, not a neuroscience finding.
- Correction: a mistaken claim is replaced, such as a corrected birthday.
- Temporal change: both statements were true at different times, such as leaving a job.
- Context: someone dislikes commercial flights but enjoys small planes.
- Perspective: two relatives describe the same trip differently.
- Unresolved conflict: the evidence cannot reconcile two accounts, so the disagreement is kept and a focused question asked when it matters.
What should a time-aware memory record contain?
A proposed record would separate source evidence, the historical state it describes, the best-supported current interpretation, and anything still uncertain. A career plan that moves from interest to interview to a declined offer forms a sequence, not four values competing for one label.
The system should distinguish when a statement was made from when it was true, and an intention from a completed event. Sources should stay attached: a member’s statement, a quotation, and an AI interpretation have different evidential roles. If source material is corrected or removed, dependent summaries should be updated too.
Should surprising statements trigger automatic updates?
Surprise can prioritize review without determining truth. A striking new message might be a joke, a quotation, or fiction. A proposed review identifies the people, events, times, and sources, classifies the mismatch, and inspects the evidence before changing the current view. Consequential ambiguity should prompt a targeted question rather than a confident guess.
AI-generated summaries need the same discipline: an inference stays an attributed inference, never something the member supposedly said. Unexpected information earns investigation; evidence determines whether revision is warranted.
How could we test whether change tracking helps?
A proposed experiment would compare three policies on synthetic histories: latest statement wins; append everything and resolve later; and preserve sources, dates, claim types, and links between earlier and current states. Histories would include corrected dates, canceled plans, evolving preferences, perspectives, unresolved contradictions, and recent jokes or quotations.
Questions would pair present and past: what is the plan now, what was it before, when did it change, and was the old statement mistaken or accurate at the time? Scores would cover current accuracy, historical fidelity, attribution, abstention, unnecessary clarification, and cost. The expectation that evidence-aware tracking performs better is a hypothesis to test, not a reported result.
What does good change tracking look like to a person?
Imagine telling an assistant last year that you lived in one city and this year that you moved. A careless system might answer with either city, or blend them. A time-aware system would say where you live now and, if asked, recall that you used to live somewhere else and roughly when that changed.
The same applies to plans, jobs, relationships, and preferences. People change, and a memory that cannot represent change ends up either stuck in the past or forgetting the past entirely. Neither serves the person well.
When a system is unsure whether a new statement replaces an old one, the most respectful choice is often to ask. A short question costs little, while silently overwriting a true fact can produce mistakes that are hard to notice later.
In one sentence
- Provenance
- The source history that shows where a claim came from and what evidence supports it.
- Temporal change
- A transition in which different claims can be accurate at different times.
- Latent cause
- A hidden situation or state inferred to explain a set of observations.
Key takeaways
- A correction, a changed plan, and a contextual difference require different handling.
- Current answers should remain linked to dated evidence and relevant historical states.
- A surprising statement should trigger review, not automatic belief.
- Evidence-aware change tracking is a design proposal whose benefits and costs need testing.