Applied AI · 4 min read
AI memory needs a way to change its mind
Persistent context becomes useful when a system can distinguish a current decision from an old fact, a tentative idea, or a superseded plan.
In August, my own memory system kept handing me things I had already retired. Notes I had replaced came back from search as if nothing had changed, and a fact that had expired stayed active and kept getting assembled into the context an agent would read. It remembered the earlier conversation perfectly. That was the problem.
The appeal of AI memory is easy to understand. Explain the project once. Stop repeating the same preferences. Pick up the work tomorrow without another orientation session. Then the project changes. A temporary workaround becomes unnecessary. A proposal gets rejected. The team changes its mind, and the assistant keeps faithfully serving the old version.
Working on ATLAS, my project-memory system for AI tools, has made me think about remembering as a maintenance problem. Saving more context only helps if there’s also a way to tell what still applies. It’s one of the more interesting problems in AI right now, partly because people have it too. We all carry a few beliefs that were true when we learned them.
A suggestion can quietly become architecture
Consider three notes: we could use a queue for this task; we decided to use a queue; the queue is deployed. They share most of their vocabulary. They describe very different states of the work.
If a system retrieves them as interchangeable evidence, a suggestion becomes a plan without anyone deciding it. A future agent may start building on a decision nobody made. The original words can be preserved exactly while the resulting guidance is wrong.
ATLAS separates memories such as decisions, observations, blockers, preferences, and rejected alternatives. That structure gives the system something to reason about beyond similarity. It still depends on the quality of the record. Calling a guess a fact doesn’t improve the guess.
Keep the reason and the limits
‘Use this service’ is less useful than a decision that explains why it was chosen and which project it applies to. A service selected for a prototype may be the wrong choice for a different workload. Without the original constraints, the memory turns a local decision into a universal preference.
I want a record to carry its source, its scope, and enough rationale to revisit it. Confidence can be helpful context, but a number doesn’t replace evidence. A direct decision from the person responsible and an inference from an old conversation should stay distinguishable.
Rejected alternatives deserve space too. Knowing that a team considered an approach and why it declined can prevent a lot of enthusiastic rediscovery. Sometimes the constraint has changed and the idea deserves another look. Preserve the reason so that judgment is possible.
Old notes don’t retire themselves
Adding a new note doesn’t necessarily retire the old one. Both may keep appearing in search, and the longer or more often repeated version may look more persuasive. The person asking the question shouldn’t have to settle that conflict from scratch every time.
ATLAS includes an explicit way to supersede a memory: keep the earlier record for history and create its replacement. It also supports validity windows for temporary facts and verification to mark a memory as still current. These are ways to maintain the record, not a guarantee that every stored statement is true.
The distinctions matter. Something old can remain correct. Something recent can be speculation. A temporary instruction may expire on a known date, while an architectural decision needs review only when its assumptions change. One blanket freshness score can’t express all of that.
Test a change of mind
A useful memory test starts with a decision, then changes it. Ask the system what applies now. Can it return the replacement, explain why the earlier choice changed, and keep an unrelated project’s preference out of the answer?
That tests something more useful than recalling an isolated fact. Repeat it with a rejected proposal, an expired workaround, and a note whose source is uncertain. Watch whether the system keeps those distinctions when it summarizes.
You can apply the same discipline to a plain Markdown project log. Mark decisions clearly, link replacements, and give temporary instructions an end condition. A small record that accurately describes the present beats a detailed archive that leaves every past direction in force.
The goal I care about is continuity: enough history to understand the work, and enough maintenance to move it forward. Forgetting well turns out to be a design problem. Remembering why we changed our minds is part of it.