The optional systems model uses an explicit directed topology:
dentate gyrus → CA3 → CA1 → cortex
Each region independently declares population size, sparsity, plasticity, and consolidation time constant. Encoding ranks neurons from context drive plus seeded heterogeneity, then creates a bounded number of directed synapses per source member. Population identity is the mean regional Jaccard similarity to encoding; synaptic identity is a separate Jaccard measure over edge endpoints.
Passive time reduces hippocampal strength and slowly raises cortical strength. Replay traverses regions forward or reverse, strengthens cells and edges, and favors cortical transfer. The queue exposes priority, capacity, and interruption. In the neurogenesis family, dentate-gyrus replacements rewire every incident edge while preserving population size.
Exact recall weights hippocampal support more strongly; the related-context probe weights cortex more strongly. The resulting precision-to-generalization curve is therefore an explicit consequence of declared regional weights and must not be interpreted as a fitted biological measurement.
The simulator is a sparse recurrent population model designed for controlled computational experiments. It intentionally keeps every state variable bounded and every random choice reproducible. The implementation is not fitted to biological units.
For neuron i and stimulus x, allocation score is:
score_i =
stimulusWeight * x_i
+ excitabilityWeight * excitability_i
- 0.25 * inhibition_i
+ seededUniformNoise_i
The top engramSize neurons become members. Stable identifier ordering resolves exact ties. Initial membership strength
combines encoding importance, stimulus drive, and excitability and is clamped to (0,1].
Encoding raises member excitability by encodingExcitabilityBoost. Between events, excitability returns exponentially
toward its neuron-specific baseline:
excitability_i(t + d) =
baseline_i + (excitability_i(t) - baseline_i) *
exp(-excitabilityRelaxationRatePerDay * d)
That explicit transient is what makes the temporal-linking experiment testable. Both the boost and relaxation rate are plan parameters; they are dimensionless model controls rather than fitted biological constants.
Related stimuli are constructed by retaining a requested fraction of strongly driven source channels and replacing the remaining active channels. Similarity therefore enters allocation through controlled input overlap rather than a special “related memory” flag.
A partial cue clamps a subset of neurons. At each iteration, memory support is the membership-weighted mean activation of its population. Neuronal recurrent input aggregates the supports of all engrams containing that neuron:
recurrent_i =
recurrentGain * mean_over_memberships(
membership(i,m) * support(m))
Global inhibition is proportional to mean population activation. Intrinsic inhibition is fixed per neuron. For an evaluated target, learned directed pair inhibition suppresses competitor-dominant neurons:
net_i = recurrent_i - globalInhibition - intrinsicInhibition_i - pairSeparation_i
candidate_i = logistic(activationGain * (net_i - activationMidpoint))
activation_i' = leak * activation_i + (1 - leak) * candidate_i
Cue activation is clamped from below on every iteration. The same equations are used for every memory; the expected identifier is used only for directed separation and evaluation metrics.
When a new memory is encoded, active-channel Jaccard similarity determines initial directed separation in both directions:
pairInhibition(a,b) = pairInhibitionRate * stimulusJaccard(a,b)
TrainInhibitorySeparation can increase one direction independently. This is an explicit model intervention for comparing
selectivity with and without inhibitory plasticity.
Memory consolidation follows a bounded exponential approach:
consolidation' =
consolidation + (1 - consolidation) *
(1 - exp(-consolidationRatePerDay * elapsedDays))
Membership decay is slower for consolidated memories:
retention =
exp(-ln(2) * elapsedDays / membershipHalfLifeDays *
(1 - 0.8 * consolidation))
Expected population turnover is:
engramSize * driftFractionPerDay * elapsedDays
The integer portion is applied and the fractional portion is sampled once. Weak members are removed; nonmembers with high original stimulus drive and current excitability replace them. Engram size is invariant.
Mechanistic allocation uses the score above. The random control removes both stimulus and excitability terms while preserving inhibition, seeded noise, population size, and stimulus sparsity; it independently shuffles the stored stimulus so the original correspondence cannot re-enter through drift. The permuted-stimulus control rotates the stimulus vector by a fixed half-population offset before scoring and stores that rotated vector for subsequent drift and similarity. This prevents the original correspondence from re-entering later dynamics.
For a requested coupling fraction, co-replay selects the strongest nonshared source members and replaces the weakest nonshared target members. The operation is symmetric, deterministic, and preserves each engram's cardinality. New membership strength is the bounded mean of the inherited target strength and source strength. The simulator exposes this as an intervention rather than applying it silently during time advancement.
Checkpoint comparison reports three distinct quantities:
population identity = Jaccard(member identifiers)
representation similarity = cosine(weighted membership vectors)
stimulus alignment = cosine(weighted membership, encoding stimulus)
None of these is treated as behavioral recall success.
- expected activation: membership-weighted target activation;
- strongest competitor activation: maximum corresponding value for another memory;
- selectivity margin: target minus strongest competitor;
- completion ratio: fraction of target members above activation threshold;
- distortion ratio: active, memory-bearing neurons outside the target divided by all active neurons;
- activation cost: mean neuronal activation across iterations;
- success: target wins and exceeds the configured success threshold.
populationIdentity = intersection(beforeMembers, afterMembers) / union(beforeMembers, afterMembers)
behavioralStability = clamp(1 - abs(beforeTargetActivation - afterTargetActivation), 0, 1)
Keeping these metrics separate is a central experimental invariant.
- Saturating activation can conceal differences between memories; inspect margins and component metrics.
- Drift, consolidation, and decay are coupled engineering rules, so ablations are required.
- Temporal-linking results depend on the configured encoding boost and relaxation timescale; report both with each run.
- Recall is evaluated against a declared target and is not a free-running cognitive decoder.
- Parameter sweeps should preserve seeds and report failures, not only successful configurations.
- Version-three checkpoints serialize options and random continuation. Legacy schemas require their original options.
- Biological interpretation requires independent empirical validation.