Computational
Formally separating updating, competitive learning, extinction and narrative reinterpretation. As long as they are conflated in the model, they will be conflated in the data.
From here on, everything is declared hypothesis. Six questions we consider testable, ordered from those that can be tackled now to those that would take a decade and permissions nobody holds today.
A hypothesis is worth something if you can picture the experiment that would refute it. For each of these we have written down what would falsify it: where that line is missing, it is not research, it is a wish with a technical name.
The order is not by importance but by distance. The first three can be started with tools that already exist. The last three depend on results we do not have yet and, in one case, on a regulatory framework that does not exist.
Each step exists to make the next one possible and to stop the project if it does not hold. Skipping one would mean offering people a protocol whose behaviour we do not know.
The path is designed to fail early and cheaply, not to reach the end.
Formally separating updating, competitive learning, extinction and narrative reinterpretation. As long as they are conflated in the model, they will be conflated in the data.
Validating cues and time windows on laboratory memories, not autobiographical ones, with preregistration and independent replication before any extension.
Multimodal diary, cognitive tasks, physiological signals and an adaptive engine that stays transparent throughout. No instrumentation beyond what a person can take off.
EEG and peripheral measures as feedback, without attributing to them any ability to read the content of a memory. It is a measure of state, not of meaning.
Only with universities, regulators and independent ethics committees. It is not a step a company can take alone, and we do not believe it should be able to.
Null results are the part of the literature that is missing today, and the reason the field is so heterogeneous. Preregistering and publishing even when it does not work is part of the method, not a gesture of transparency.
The implant currently in clinical study at Neuralink records signals from the motor cortex so that people with paralysis can control external devices. It does not decode autobiographical memories, it does not operate on the hippocampus, it does not insert memories. Confusing that device with a memory technology is the most common misreading of this subject.
Any future integration would require bidirectional interfaces and different neural sites: reading biomarkers of reactivation and lability, estimating cognitive state, and possibly closed-loop stimulation synchronised to the millisecond. There are findings showing that stimulation synchronised with sleep spindles can influence consolidation, and equally that imprecise timing can degrade it. That symmetry is the thing to keep in mind: the same lever that might improve can also worsen, and the difference lies in a precision we do not have outside the laboratory today.
The invasive level would therefore remain clinical, experimental and subject to independent protocols. It is not a stage on our path: it is a scenario we describe because others describe it badly.
Attenuating a response is not changing a person, but that difference has to be defended, not assumed. A memory is part of who we are, and the reaction it carries is sometimes the very part a person wants to keep. The objective is chosen by them, not by the protocol, and that choice must be withdrawable while the process is under way.
Consent here has a problem of form. Someone accepting an intervention on their own reactivity cannot judge the outcome from where they stand: they judge it afterwards, and with that reactivity already changed. Consent therefore cannot be a form signed once; it has to be renewed in stages, with a real possibility of stopping and with no cost attached to stopping.
The harm is asymmetric. A protocol that does not work leaves things as they were. A protocol that induces a false memory, or that dampens a response that was needed, does harm that cannot be undone. That asymmetry requires every design choice, in case of doubt, to fall on the side of not intervening.
No application outside consent. We rule out covert grafting, coercion, deceptive alterations of memory and any use on people who have not freely chosen to take part. We also rule out the uses somebody would call benevolent and that were never asked for by the person subjected to them.
References: Nader, Schafe & LeDoux, 2000 · Hupbach et al., 2007 · Kindt, Soeter & Vervliet, 2009 · Schiller et al., 2010 · Hardwicke et al., 2016 · Elsey, Van Ast & Kindt, 2018 · Sinclair & Barense, 2018 · Geva-Sagiv et al., 2023 · Neuralink PRIME Study
If the labile window could not be estimated from the outside, if updating propagated uncontrollably, or if every effect vanished within a month, MNEMOS would have no path. Those are possible outcomes and we treat them as such.
In that case the useful result would be having established it with a method others can reuse, and having said so. A research programme that cannot fail is not a research programme.
MNEMOS only makes sense inside a network: universities, cognitive neuroscience groups, ethics committees. If you work on reconsolidation, memory modelling or causal inference in longitudinal settings, write to us.
Let’s talk →