MNEMOS™ · Progress

Where we
currently stand.

The programme runs in phases, and every phase has a gate: a criterion written down before the work starts, which decides whether we move on or stop. This page says where we are today, what we are doing right now, and what has not been done yet.

Phase 1 of 4 · under wayNo participants involvedEntirely in simulationUpdated 5 September 2026

Four phases,
four gates.

The phases are not a calendar, they are a sequence of conditions. Each one produces the evidence required to enter the next, and each one can end in a negative result: in that case the result is published anyway, and the programme either stops or is reformulated.

Every gate criterion is preregistered, that is, filed publicly before the data it will judge has been collected. It is the only condition under which a threshold means anything: decided afterwards, it always moves by exactly as much as it takes to clear it.

  1. PHASE 1Under way

    Computational model

    A simulator of memory dynamics, entirely in silico. It narrows the hypothesis space and sizes the next phase. No headset, no sensors, no participants.

  2. PHASE 2Planned

    Synthetic paradigms

    Laboratory material, neutral and non-autobiographical. The measurement chain is validated where the ground truth is known by construction: if it does not hold there, there is no reason to trust it anywhere else.

  3. PHASE 3Planned

    Full loop with participants

    Autobiographical material, with consent and with human authorisation for every scene. Four conditions compared, passive viewing among them, and three families of measure kept apart.

  4. PHASE 4Planned

    Closed loop, persistence, propagation

    Real-time feedback, follow-ups weeks apart, and the measure that decides everything: whether the effect stays on the episode or extends to neighbouring memories that were never reactivated.

Beyond the fourth phase lies clinical research with neural interfaces, which remains workable only with universities, regulators and independent ethics committees, and only where depth electrodes are already in place for clinical reasons. It is not a step a company can take on its own, and we do not believe it should be able to: the reasoning is set out in the research directions.

PHASE 1 · UNDER WAY

Simulated people
first,
then people.

Phase 1 builds a simulator of memory dynamics and involves nothing physical: no headset, no sensors, no participants, no real autobiographical data. It is entirely in simulation, and that is not a fallback: it is how you get to be wrong at zero cost.

It serves one precise purpose. At least four different processes can produce the same observable signal when a constructed experience meets a memory: the existing trace is modified; a second trace forms and competes with the first; an inhibitory association is added; or the content stays intact and only the source it is attributed to changes. As long as they are conflated in the model, they will be conflated in the data, and no result from the later phases would be interpretable.

The fourth possibility deserves an extra line, because an immersive experience is precisely what makes it more likely: a system like this one may rewrite nothing at all and merely shift source attribution. That would be a more modest and far more defensible claim, and it is better found out now than in two years' time.

01 / PARAMETERS

A narrowed space

The parameter ranges compatible with what the literature has already measured. Everything else is ruled out beforehand, not after looking at the results.

02 / POWER

How much data is actually needed

The number of trials per person the next phase has to collect for the measurement to hold. Without that number, Phase 2 is sized by intuition and you find out too late that you measured noise.

03 / FAILURE MODES

The ways of being wrong, anticipated

The ways the loop can converge on something that looks like success and is not. Finding them in simulation costs a week; finding them with participants costs the programme.

04 / SAFETY

The propagation curve

The conditions under which an update stays on a single episode and those under which it spreads to neighbouring memories. This is where a numerical stopping threshold comes from, rather than an impression.

05 / HETEROGENEITY

Differences between people

How much between-subject variance the model has to generate for personalisation to be worth anything. If it generates none, there is nothing to personalise and the programme's central claim falls at home.

06 / PREREGISTRATION

Phase 2, written first

Protocol, hypotheses, controls and gate thresholds for the next phase, filed publicly before a single data point is collected.

The Phase 1 gate, stated as a number

The model has to reproduce at least six of eight already published experimental targets, chosen and declared in advance, using a single population distribution of parameters and no refitting target by target. Two of the eight are held out of the fitting and used for validation only.

Two clarifications matter more than the number. First: what must fall inside the published interval is not only the mean effect, but also how much that effect varies between people. A model that nails every mean with near-zero variance has failed, because the programme's whole claim is that these effects are heterogeneous and that personalisation is the lever. Second: the targets include failed replications. A model that cannot produce the null results is not a model of the literature, it is a model of the convenient half of it.

Below six out of eight the model is revised, and we do not proceed to the phase involving participants.

What we are doing
right now.

  1. The priors sheet The initial parameter values and ranges, fixed and timestamped before looking at the results the model will have to reproduce. A model with many free parameters, left to itself, fits anything: this sheet is what stops us discovering after the fact that we were right all along.
  2. The target list The eight published results with their confidence intervals, and an explicit note of which two are held out of the fitting. This too comes before any code.
  3. The simulator The core of the model, with the four update routes formally separated. Versioned code and seeds, one configuration file per experiment, a report reproducible with a single command.
  4. The propagation calculation The region in which an update stays local and the region in which it spreads to neighbouring memories. It is the programme's safety endpoint, and it is computed before a participant exists, not after.
  5. The observation model The translation from internal states to the measures the later phases will actually record, deliberately contaminated with the real confounds. The same analysis pipeline that will run on real data runs on this synthetic data: that is how you learn how much data is needed, before buying the instruments.
SCOPE OF THE CURRENT PHASE

Phase 1 does not touch people. The first protocol involving participants is Phase 2, on neutral laboratory material, and it does not begin without public preregistration, a legal basis for biometric and health data, an impact assessment, and local encrypted storage.

Before the phase with autobiographical material we need a structure we do not yet have. An independent clinician for eligibility and adverse events, who cannot be the person running the session. A safety monitor with actual authority to halt the programme, not an advisory role. A data protection officer, a methodologist for preregistrations and analyses, and a supervision operator distinct from whoever built the person's model. Whoever builds does not authorise, whoever runs the session does not screen, whoever analyses has no stake in the outcome: it is that constraint that makes the result inspectable by a third party, and it is one of the reasons Phase 1 comes first.

This page describes the method and the criteria, not the implementation. The system is the subject of patent application no. 102026000024559 of 24 August 2026, and the technical matter the phases produce stays confidential until its respective filings. What we publish are the conditions we have set ourselves and the results, negative ones included.

Continue in The directions · The intelligence layer · Back to MNEMOS

What we have
not done
yet.

A status page is worth little if it only lists what went well. Below are the four things that, as of today, have not been done, and that no result from the current phase may treat as settled.

NOT YET / 01

No participants

No person has been involved in any form and no biometric or health data has been collected. The first collection comes in Phase 2, on laboratory material and after a public preregistration.

NOT YET / 02

No validated familiarity index

That familiarity can be estimated from bodily signals better than chance is the hypothesis Phase 2 has to test, against a matched protocol that changes parameters at random and against a decoder built from image statistics alone. Today it is not demonstrated, and it may turn out to be false.

NOT YET / 03

The loop has never run on autobiographical material

The patent application describes the architecture, and an architecture can be shown. Whether a constructed experience, once walked through, leaves an encoding signature distinguishable from passive viewing is an open question, and it is addressed in Phase 3.

NOT YET / 04

Propagation has not been measured

And it is the endpoint that decides whether the programme continues. What has to be shown is that the effect does not extend beyond a margin declared in advance, not that “no effect was found”: those are two different statements, and with small samples the second one is almost always obtained and proves nothing.

Log.

The dates that matter, most recent first. This page is updated when a phase changes state or when a gate opens or closes, not on a fixed schedule: a log that updates every month whatever happens stops saying anything.

  1. Phase 1 opens

    Version 1.0 of the detailed four-phase plan closed: scope, measures, controls, gates and stopping criteria for all of them, written before starting the first. The priors sheet and the target list are in progress, and both come before the simulator code.

  2. MNEMOS™ trademark application

    Number 302026000148432.

  3. Patent application for an industrial invention

    Number 102026000024559, on the closed-loop system. The ANALYTIKO™ trademark application no. 302026000148294 was filed the same day.

  4. The programme goes public

    Seven hypotheses, each with its own falsification condition, the five-step path and the ethical boundary. From that day “null results get published” is not an internal intention: it is a written commitment anyone can check.

Phase 1 can be done
alone.
The others cannot.

The phases involving participants require an independent clinician, a safety monitor with authority to halt, a methodologist and an ethics committee. Analytiko™ is looking for collaborations with universities and cognitive neuroscience groups for Phases 2 and 3, and is glad to put the computational model in front of independent review before it drives a single decision.

Propose a scientific collaboration →