Research

Facts Are the Cheap Half of Memory

Every AI memory product stores what happened. None of them store what you corrected.

By Merryl DMello, DMello Advisory · 4 min read

The claim: the AI memory products on the market, every one of them, store facts: preferences, project context, decisions. Facts are the cheap half. The valuable half is behavioral: which corrections recur, for which models, in which situations, and which rules actually changed anything. That half is not stored by any product. It is sitting, right now, in the correction records of every heavy AI user, and one fully mapped corpus shows both the gap and what closing it looks like: 40+ dated incident files recording corrected behaviors, recurrence data proving that written rules did not bind, and one subsystem, the only one in the entire stack, that earns its rules through measurement.

The gap, mapped

The corpus behind this paper contains a personal instruction stack of unusual discipline: more than 100 rules, every one with provenance. Forty-plus dated incident memory files across 25 projects, each written at the moment of friction, each recording exactly one corrected behavior. If behavioral memory existed anywhere, it would exist here.

It does not, and the recurrence data proves it:

• One model repeated wrong assumptions about external state across three separate projects, despite a rule against exactly that.
• Another model's premature handoffs recurred three times inside one project, again against an existing rule.
• A third's stale-external-state pattern recurred twice in one project.

A rule the same model violates three times in a month is not load-bearing. The stack's promotion loop, from incident file to global rule, is entirely manual, and the corpus documents open loops: an incident marked "permanent fix not yet done" was still open months later. The friction was recorded, filed, and then re-experienced.

The market's answer to this is more storage: memory products that remember your preferences with increasing fidelity. But preference recall does not address the recorded failures above. In every case the rule already existed and was already stored. The system knew, and the model did it anyway.

The one subsystem that works

There is exactly one exception in this stack, and it works because it treats rules as hypotheses rather than possessions. Skills enter as candidates, run as experiments, and get scored: mutate, test, measure, then promote on statistics with a human veto, or discard. That subsystem has run 430 tests. It is the only part of the stack where a rule's right to exist is earned by a measured effect, and, not coincidentally, the only part where dead rules actually die.

The contrast with the other 100+ rules is the paper's argument in miniature. Storage plus sentiment keeps every rule forever, including the ones that never worked and the ones that work but are never checked. Measurement plus promotion produces a stack that converges on the rules that bind.

What closing the gap looks like

The correction record itself is the raw material, and it is richer than most teams realize, because it includes the channel nobody reads: interrupts (131 of them in this corpus, against roughly 51 typed corrections). A behavioral memory that works would do three things:

1. Detect recurrence mechanically: same correction type, same model family, same situation class, across projects and months. The 180-event hand-checked taxonomy in this corpus shows the classes are stable enough to count.
2. Auto-draft the candidate rule, with provenance: which incidents, which dates, what delta it is supposed to cause.
3. Promote or kill on evidence: compare matched sessions with and without the rule, dated cohorts, and delete the rule when the delta is zero. The natural experiments already exist in any long-running corpus; months of the same user, same projects, same models, with and without a given skill loaded.

Step 3 is what no product does and what separates a memory from a diary. A rule that cannot lose its job is not knowledge. It is clutter with tenure.

What would change my mind

• A shipping memory product that stores correction recurrence per model and situation, not just facts and preferences. I can find none; the category markets recall, not behavior.
• Evidence that stored facts reduce correction rates. Every measurement I know of, including the 288-run context-file ablation finding no measurable correctness gain, points the other way.
• A rule stack, any stack, where per-rule effects are measured and zero-effect rules get deleted. The hypothesis-testing subsystem above is the only one I have found, in one private corpus, built by hand.

The author runs dmello-advisory and builds measured AI stacks for small and mid-size companies: rules that carry their own evidence, memory that learns from friction instead of hoarding facts. Eighteen years building and leading products at Intel, HP, Rackspace, BioNTech, and Phoenix Contact. dmello-advisory.com.

All research papers

Last updated: 16 October 2026