Predictive Memory Architecture (PMA): the six components of the second wave

Added on 11 October 2026. The six PMA components are described in the paper and ran in the ZenAI production system; they are not part of the open-source packages. Our correction notice of 18 September 2026 corrects the stress ablation: under stress, 10 of 15 components reach the printed criterion in the simulation, nine with a loss and PriorityMap with a gain. The figure for TripleCopy strength at 30 days and the transition times are removed; they did not follow from the printed formulas.

In March 2026 we had 9 foundational algorithms — vmPFC-FSRS, Two-Factor Synaptic, Sim-Selection Sleep, Bayesian Confidence, and five others. They structured what gets stored, where, and with what importance.

What they did not regulate: when a memory becomes labile, which update strategy a retrieval triggers, how the system reacts under bias pressure, and which memory deserves protection from overwriting right now.

That is the job of the Predictive Memory Architecture (PMA) — a second wave of six components formally described in the ZenBrain paper. They sit on top of the 9 foundational algorithms and govern the lifecycle of memory over time.

Here is a compact explanation of the six.

1. NeuromodulatorEngine — four channels, tonic + phasic

The brain has no global learning rate. It has a whole toolkit of neuromodulators that act differently in different places depending on the situation. The NeuromodulatorEngine emulates four of them:

• Dopamine (DA) — VTA, exploration and novelty bias
• Norepinephrine (NE) — Locus Coeruleus, learning rate and attention to prediction error
• Serotonin (5-HT) — Raphe, consolidation patience
• Acetylcholine (ACh) — Basal Forebrain, attention ratio between old and new information

Each channel has a tonic baseline (b = 0.5) with homeostatic drift (τ_decay = 0.95) and phasic bursts on events (5-minute half-life). DA and 5-HT are linked by an opposition coupling (−0.3) — a realization of the serotonin-dopamine balance well-documented in reward processing.

The engine outputs four modulation parameters consumed by other PMA components: learning rate (NE-driven), exploration bias (DA-driven), consolidation patience (5-HT-driven), attention ratio (ACh-driven).

In the paper experiments the engine produces a mean tonic drift of 0.469 (6.2 % from baseline) over 1,000 simulated events — homeostatic stability is preserved — and a DA/5-HT correlation of −0.130 (p < 0.01), which validates the opposition coupling.

2. ReconsolidationEngine — four update modes, not one

The classical view of memory: stored once, then read, then maybe overwritten. The neuroscience reality is subtler. When a memory is retrieved, it becomes labile for about 10 minutes — it can be modified or erased — and then re-stabilizes in a possibly altered form. This is reconsolidation (Nader 2000, Schiller 2010).

The ReconsolidationEngine implements this with four modes, gated by the effective prediction error PE_eff:

mode(PE_eff) =
    confirmed         if PE_eff < 0.1
    selective_edit    if 0.1 ≤ PE_eff < 0.3
    integration       if 0.3 ≤ PE_eff < 0.7
    new_episode       if PE_eff ≥ 0.7

The effective PE is neuromodulator-gated: PE_eff = PE_raw · (1 + 0.3·NE − 0.2·5HT). High NE (stress, attention) amplifies PE; high 5-HT (patience, stability) dampens it.

Every reconsolidation event is logged with an original snapshot — rollback is possible if the update later turns out to be a mistake. In the paper experiments the PE-to-update-mode classification reaches ≥ 95 % accuracy with correct contradiction detection in 100 % of test cases (precision = 1.0).

3. TripleCopyMemory — three traces, three time constants

Probably the most surprising component. Instead of storing a memory once, TripleCopyMemory stores it three times — with divergent decay dynamics:

S_fast(t) = S₀ · e^(−t/τ_f),       τ_f = 4 h
S_med(t)  = 0.8 · S₀ · e^(−t/τ_m), τ_m = 14 d
S_deep(t) = S₀ · log(1 + t/τ_d),   τ_d = 7 d

• The fast copy delivers vivid immediate access that fades within hours.
• The medium copy persists across sessions with standard exponential decay.
• The deep copy grows logarithmically — it encodes the compressed essence and gets stronger over time.

The effective strength is S(t) = max(S_fast, S_med, S_deep). That produces a characteristic dominance transition: as a memory ages, the lead passes from the fast copy to the medium and then to the deep copy.

This matches the systems consolidation theory in neuroscience: detail-rich episodes fade, compressed gist representations survive.

4. PriorityMap — four-dimensional, with amygdala fast-path

Memory without prioritization gets buried in noise. The PriorityMap computes a four-dimensional score:

P = w_s · s + w_e · |v| + w_r · r + w_g · g

with saliency s, emotional valence v, reward relevance r, goal alignment g, and default weights (w_s, w_e, w_r, w_g) = (0.2, 0.25, 0.25, 0.3).

The decisive mechanism is the amygdala fast-path: at emotional intensity |v| > 0.6 the system guarantees P ≥ 0.5, regardless of the other dimensions. This mirrors McGaugh's (2004) finding that emotionally charged memories receive preferred consolidation — even when not task-relevant.

The weights are dynamically adjusted by neuromodulator state: DA amplifies saliency, NE amplifies emotion, ACh amplifies reward, 5-HT amplifies goal alignment. The same PriorityMap scores can produce different values in different states, reproducing biological behavior.

On a synthetic benchmark with 50 items and ground-truth importance the PriorityMap reaches NDCG@10 = 0.997 (vs. 0.680 chronological, +46.6 %).

5. StabilityProtector — protection for mature memories

Reconsolidation is powerful — but it also opens an attack window. What if a single false statement could overwrite a memory established for years?

The StabilityProtector prevents this with a lock score L and a rigidity factor ρ:

L      = 0.3 · log₂(1+a) / log₂(11)
       + 0.3 · c
       + 0.2 · min(d/365, 1)
       + 0.2 · is_core
ρ      = 1 + 0.1 · log₂(1+d)
update ⇔ PE ≥ 0.5 + 0.3 · L · ρ

with access count a, confidence c, age in days d, indicator is_core for core memory. Mature memories (a large, c large, d large, or core status) get high lock scores and rigidity factors — an update demand with low PE is rejected. Genuinely new information with high PE penetrates the protection.

This is analogous to NogoA receptor signaling and HDAC3 epigenetic regulation: the biological brain actively protects mature circuits against casual rewriting.

6. MetacognitiveMonitor — when the system notices it's biased

The final component watches the system itself. It tracks three bias types:

• Confirmation bias — asymmetric acceptance of positive vs. negative evidence
• Recency bias — overweighting recent memories
• Retrieval asymmetry — certain domains systematically under-retrieved

It opens novelty windows (10 minutes) after high-PE events (> 0.7) where encoding is temporarily boosted — and generates calibration alerts when systematic biases exceed a 30-day threshold. Efficiency tracking over the rolling 30-day window produced badges that surfaced in the ZenAI UI — closing the feedback loop.

In the paper experiments confirmation-bias detection reaches precision = 0.832 and recall = 0.975 across 50 synthetic scenarios per seed. Urgency keyword detection produces 0 false negatives on German and English test phrases.

How the six work together

PMA is not a stack of six independent modules. It is a coupled system:

• The NeuromodulatorEngine delivers the global tonic values all others read.
• The ReconsolidationEngine relies on NE / 5-HT values to modulate PE_eff.
• The PriorityMap reads the same tonic values to dynamically adjust weights.
• The StabilityProtector becomes stricter when 5-HT is high (more patience) and more permissive when NE is high (more learning readiness).
• The MetacognitiveMonitor observes all four and triggers novelty windows that briefly boost learning rate and encoding strength.
• TripleCopyMemory is the only component that runs independently — it is the substrate layer the other five operate on.

The coupling shows up empirically in the paper's ablation: under moderate conditions (decay = 0.15/day, 45 days) all 6 PMA components are cooperatively redundant (ΔQ ≤ 0.1 % when removed individually). But removing all six at once — testing "Foundational-only" against "Full" — collapses the system by −67.5 %. PMA is a resilience backbone, not an optional add-on.

Under stress (60 days, decay = 0.25/day) the coupling becomes visible: without NeuromodulatorEngine quality drops by 83.0 %, without TripleCopyMemory by 93.7 %, while iMAD and MetacogMonitor deliver their contribution in ranking precision rather than retention rate. Per our correction notice of 18 September 2026, PriorityMap also reaches the printed criterion under stress, though with a gain (10 of 15, nine of them with a loss).

What this means in practice

The 9 foundational algorithms were architecture: what to store, where, how to structure. The 6 PMA components are dynamics: when to become labile, which update mode, when to protect, when to boost.

You need both. Either one without the other collapses.

Read more

• Predecessor post: How we built memory into an AI OS — the 9 neuroscience algorithms
• Cost ratio: 91 % of the accuracy at 1 % of the tokens
• Paper: ZenBrain on Zenodo
• Code of the algorithm library: github.com/zensation-ai/zenbrain; the six PMA components are described in the paper and are not part of the open-source packages.