The most misunderstood part of the project. No technology reads the content of a memory: here AI serves to choose what to present, and when, and to work out whether anything genuinely changed.
Stating the limits first is not modesty. It is what makes everything else testable: if the perimeter is clear, every claim inside it can be put to the test.
Four things you will hear attributed to systems like this one, and which are true neither today nor in any reasonably foreseeable future version.
No technology decodes the content of an autobiographical memory. The neural interfaces currently in clinical study read motor intentions from the motor cortex, which is a different signal in a different area. The distance is not one of degree, it is one of kind.
The model knows what the person has recounted, not what happened. Treating internal coherence as proof of truth would be a serious methodological error, and in a sensitive context a harm as well.
The objective of the intervention is chosen by the person. A system optimising towards an affective state decided by others would not be memory research: it would be a conditioning technology, and that is precisely what we rule out.
No stage is designed to operate without competent human oversight. The engine proposes, a person decides, and the decision remains open to inspection afterwards.
The literature on reconsolidation is not stalled for want of a theory. It is stalled because the conditions that open the window are narrow, individual and not observable from the outside.
Which signal to use for reactivation. How much mismatch to introduce against expectation. How old that memory is, how strong, how many times it has already been recalled. In published studies these parameters are set with coarse heuristics, identical for every participant, and then the heterogeneity of the results comes as a surprise.
This is exactly the shape of problem where an individual statistical model has something to say. Not because it is intelligent: because it can hold more variables per person than any hand-written protocol can carry.
Estimating, for this person and this memory, which recall produces enough surprise to destabilise without forming a new memory alongside it. Today this is the worst-controlled variable in the entire literature.
It cannot be observed directly. What can be observed is response latency, hesitation, the structure of language, physiological signals: from those, a probabilistic estimate of the state of the system, explicitly uncertain.
Updating, extinction, habituation, placebo, compliance with the experimenter, regression to the mean. Without separating them, any positive result is indistinguishable from an artefact. It is the least spectacular part and the most decisive.
A computational model of memory makes it possible to explore thousands of combinations of cues, timings and content before proposing one to a person. It shrinks the search space and moves the risk from human beings to computation.
That studies disagree may not be noise. There may be subgroups with genuinely different dynamics. Finding them requires individual longitudinal data and models that do not average everything away at the first step.
Every cue proposed, every piece of content integrated, every decision made by the engine is recorded and attributable. Without this there is neither scientific review nor genuine informed consent.
The representation we build is a temporal graph of people, places, episodes, stated emotions and contradictions, fed only by what the person has confirmed.
It is not a copy of their memory, and it matters not to call it one even as shorthand. It is a model of how that person recounts their memory at a given moment, with all the omissions and rewritings the account carries with it.
That makes it less fascinating and far more useful: contradictions are not errors to be corrected, they are signal. They are the points where the account is unstable, and therefore the natural candidates for understanding where any margin exists.
An adaptive engine selecting cues and content within predefined limits is technically a constrained reinforcement learning system. The engineering temptation is to close the loop and let it optimise. That is the one thing we will not do. Not out of formal caution, but because the objective to be optimised, in this domain, cannot be measured without asking the person living it. A system maximising an affective indicator would find shortcuts nobody would want, and would find them faster than we would notice.
In this domain, words do damage. “Rewriting”, “erasing” and “grafting” do not describe capabilities available today. The controlled grafting of new memory traces is nonetheless the declared frontier of MNEMOS: reactivating, integrating, restabilising and measuring are the observable processes from which to start in understanding whether it could become possible. The data will decide how far this vocabulary can translate into a real capability.
Continues in The directions · Back to MNEMOS
It is establishing that an observed change really is the effect you were looking for. The directions of exploration, the translational path and the ethical boundary are on the next page.
The directions →