Steve HutchinsonBig Pines
·7 min read·Stage 18·Cognitive Substrate

Forgetting System

This article describes the forgetting system that suppresses, compresses, retires, and prunes memory so cognition remains usable over time.

Forgetting as executive control

Forgetting engine: the memory graph is evaluated against retrieval feedback and contradiction records and dispatched to suppress, compress, retire, or prune.
Forgetting engine: the memory graph is evaluated against retrieval feedback and contradiction records and dispatched to suppress, compress, retire, or prune.

Memory growth without forgetting produces noise. Every retained trace competes for retrieval, attention, and consolidation. As the memory index grows, retrieval quality degrades unless the system actively manages what stays and what goes.

But forgetting is not just about managing index size. Arriving at stage 17 in the architecture's development, after the system has accumulated policy drift, identity pressure, attention constraints, temporal context, and budget limits, the question of forgetting is qualitatively different from what it would have been at the start. Forgetting is now executive control over an active memory system: which traces remain retrievable, which contradictions should retire from active belief, and which repeated details should become abstractions.

Why decay causes catastrophic convergence

Experiment 13 ran a deliberately alarming test: what happens to memory rankings if the system accumulates good Hebbian compounding for 100 turns, then decays without any further reinforcement?

The results were stark. After reinforcement, cluster-A (trusted, high-importance memories) had average retrieval priority 0.748. Cluster-C (contradictory, low-importance memories) had average retrieval priority 0.238. The gap was 0.510.

After 50 epochs of decay with no re-consolidation, cluster-A fell to 0.232 and cluster-C fell to 0.155. The gap was only 0.077. After 100 epochs, cluster-C (0.100) actually exceeded cluster-A (0.072). The system had inverted: contradictory memories were now more likely to be retrieved than trusted ones.

The mechanism is multiplicative decay. High-priority memories start high and fall fast in absolute terms. Low-priority memories start low and fall slowly. Eventually they converge, and at very long time horizons they invert because the floor is higher relative to the starting value for low-priority memories.

This is biological long-term potentiation behavior: synaptic connections that are not re-activated eventually decay to baseline regardless of how strong they once were. The implication for cognitive systems is the same: memories that are never re-accessed fade, even if they were once strongly reinforced.

Re-consolidation as the fix

Experiment 16 tested whether periodic re-consolidation (a small positive boost to memories above a quality threshold) could prevent catastrophic convergence. Three conditions were compared over 100 decay epochs:

  • No re-consolidation: the gap inverted to -0.024 at epoch 100.
  • Re-consolidation every 10 epochs: gap was -0.015 at epoch 100. Reduced inversion but did not prevent it.
  • Re-consolidation every 5 epochs: gap was +0.012 at epoch 100. Inversion fully prevented.

The re-consolidation rule: every RR epochs, all memories with retrieval priority above the re-consolidation threshold (0.450.45 in this experiment) receive a small positive boost (4%4\% of importanceScoreimportanceScore). Memories below the threshold do not qualify.

This selectivity is the key insight. Re-consolidation is not a uniform memory refresh. It targets trusted memories (those above the quality threshold) and selectively extends their advantage. Cluster-C memories, with low retrieval priority, fall below the threshold and receive no boost. The mechanism is therefore self-reinforcing in the right direction: good memories receive re-consolidation, bad memories do not.

The minimum effective interval (5 epochs at 4% boost for this corpus) depends on the decay rate and the quality distribution. Faster decay or a more extreme quality gap would require more frequent re-consolidation. The principle is general even when the exact parameters are corpus-specific.

Severity-stratified forgetting

Experiment 19 demonstrated how the forgetting system handles memories with different importance levels. Four incident windows from the operational signal experiments (normal, degraded, outage, recovery) were subjected to the forgetting plan.

At baseline age (0 days), the results were clear. All 50 outage signals (importance 0.84 to 0.96, retention score above 0.584) were retained. All 60 degraded signals were retained. 35% of normal signals and 30% of recovery signals were suppressed, because their low importance scores (0.14 to 0.30) produced low retention scores.

At 60 days, the picture changed further: 0% of normal signals retained, with 26 pruned and 14 suppressed. Retention score at 60 days was approximately 0.21, below the suppression threshold of 0.35.

One finding deserves emphasis: the compress action was not triggered for any normal-window signal at 60 days, even though compression (which consolidates memories worth preserving into abstractions) is the natural action for aging important memories. The reason is that compression is gated by retentionScore > suppressionThreshold. Low-importance signals hit the suppression threshold (0.35) before reaching the compression branch. Compression is reserved for memories worth consolidating; it does not fire for memories that are already below the quality bar.

This is not a bug; it is the correct architecture. Compressing low-quality memories into abstractions would propagate low-quality content into a more permanent form. The forgetting system correctly sends low-importance aging memories to suppression and prune, not to compression.

The four forgetting actions

The forgetting system uses four distinct actions, each appropriate for different situations:

Suppress: remove from ordinary retrieval without deleting the record. The memory still exists and can be accessed by explicit ID or during reflection, but it will not surface in normal context hydration. Useful for stale plans, temporarily irrelevant evidence, or memories that distract from better evidence without being incorrect.

Compress: consolidate into a higher-level abstraction. Rather than losing the information, the system promotes it to a more general representation. A sequence of similar incident memories might compress into a pattern memory that captures the common structure without preserving the individual details. Compression connects forgetting to the abstraction system.

Retire: mark as superseded by later evidence. When newer memories explicitly contradict earlier ones, the earlier memories can be retired from active belief. Retirement preserves the historical record (the memory still exists with a retired status) while preventing outdated beliefs from steering future action. This is different from deletion: the audit trail is intact.

Prune: remove graph edges below a quality threshold. Memory associations that are weak or misleading degrade retrieval quality by pulling in irrelevant neighbors. Graph pruning removes these edges without touching the memory records themselves. Experiment 19 showed 5 links input and 3 retained (strength above 0.15), with 2 pruned (strength below 0.15).

Why contradiction retirement is different from deletion

An important asymmetry: retirement preserves the ability to reconstruct the history of belief change. A system that simply deletes contradicted memories cannot explain why it stopped believing something. A system that retires them with a timestamp and a pointer to the contradicting evidence can.

This matters for audit and for trust. When a system's belief changes, operators and users need to be able to understand why. Deletion says "we no longer believe this, no further information available." Retirement says "we believed this until X, which showed Y, so the prior belief was superseded." The second form supports the kind of explainability that makes an adaptive system trustworthy.

A known production gap

Experiment 16 establishes that periodic re-consolidation is required to prevent catastrophic convergence: at a 5-epoch interval the trusted-memory advantage is preserved; at 10 epochs it is not. The mechanism is implemented and validated. The background job that runs it on a schedule is not yet wired in the hosted deployment. A system operating without periodic re-consolidation will experience gradual priority degradation of trusted memories over time, eventually reaching the inversion described in Experiment 13. The failure modes article covers this and the other known boundary conditions of the system in full.

Related Articles

This site collects anonymous usage data to understand how people read and navigate the blog. Accepting enables persistent reader preferences across visits.