The forty-four experiments in this series were not all confirmatory. Several were designed to find where the system breaks, and they succeeded. This article documents those failure modes - when they occur, what the mechanism is, and what the known mitigation is. None of these are surprises; they are boundary conditions of a system that behaves predictably within its designed operating envelope.
Catastrophic convergence without re-consolidation
What happens: Trusted memories invert with contradictory ones after sustained decay without reinforcement.
Mechanism: Multiplicative decay causes high-priority memories to fall faster in absolute terms than low-priority ones. Given enough time without retrieval or reinforcement, the two populations converge. At very long time horizons, they invert: the memories that once had high retrieval priority end up below the memories that were always low quality.
Experiment 13 ran this scenario deliberately. After 100 Hebbian compounding turns, cluster-A (trusted, high-importance) had average retrieval priority 0.748 and cluster-C (contradictory, low-importance) had 0.238 - a gap of 0.510. After 100 decay epochs with no re-consolidation, cluster-A fell to 0.072 and cluster-C to 0.100. The ranking inverted.
Mitigation: Periodic re-consolidation. Experiment 16 tested three conditions over 100 decay epochs: no re-consolidation (inversion at epoch 100), re-consolidation every 10 epochs (reduced inversion, not prevented), and re-consolidation every 5 epochs (inversion fully prevented, gap +0.012 at epoch 100). The mechanism: every epochs, memories above a quality threshold receive a small boost proportional to their importance score. Memories below the threshold receive nothing, so the selectivity maintains the quality ordering.
Production gap: The re-consolidation 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 this job will experience gradual priority degradation of trusted memories over time, eventually reaching the inversion state above.
Graph slot contamination
What happens: Expanding retrieval through memory graph links introduces low-quality memories at high rates.
Mechanism: The retrieval engine can allocate one or more result slots to graph neighbors of top-k seeds - memories connected via supports, contradicts, extends, or associates links. The intent is to surface related memories that scored below the top-k threshold. The problem is that contradicts links connect trusted memories to their contradictions. Expanding through those links surfaces the contradictory memories as retrieval candidates.
Experiment 5 tested one-hop graph expansion. It did not improve hit rate for the target cluster (cluster B was already reachable via session-relative novelty). More concerning, the contradicts links surfaced cluster-C memories at a 30% increased rate with no corresponding benefit. Experiment 6 tested a diversity slot that guaranteed the best graph neighbor - it improved cluster-B access but introduced cluster-C contamination at 50%.
Mitigation: Filter graph expansion by edge type. Links that support or relate memories can enrich context; links that record contradiction belong in reflection, not retrieval. The diversity slot itself is not in production use; the filtering logic is implemented but the slot is disabled pending further validation.
Parameter distribution sensitivity
What happens: Attention weights and reinforcement parameters calibrated on the synthetic corpus may underperform on a different event distribution.
Mechanism: The attention weight of 0.30 for novelty was calibrated against a corpus where a cluster of moderate-importance memories (cluster B, importance 0.55–0.65) needed to compete with a dominant cluster (cluster A, importance 0.70–0.85). At novelty weight 0.14, cluster B was invisible at all tested novelty-score values; at 0.30, it became retrievable at appropriate policy settings. The count bonus of 0.02 was calibrated against a corpus where contradictory memories had importance scores of 0.15–0.30 and trusted memories had 0.70–0.85. Both choices reflect properties of the design corpus, not universal constants.
A deployment with a different importance distribution - where all memories cluster tightly near 0.5, or where the importance gap between trusted and contradictory memories is smaller - would find different optimal values. A deployment with much higher event volume and faster novelty dynamics might require a lower novelty weight to avoid thrash.
Mitigation: Re-calibration when deploying to a new event distribution. The experiment scripts are reproducible; running Experiments 3 and 6 against a corpus sampled from the target distribution will identify the appropriate novelty weight and count bonus for that context. Parameter choices are documented with experiment citations in the code so the calibration rationale is traceable.
Silent misconfiguration
What happens: Index name mismatches return zero results without raising an error.
Mechanism: OpenSearch queries against a non-existent index name return an empty result set, not an error. If the retrieval engine is configured with an incorrect memoryIndex value - a typo, a stale environment variable, a wrong deployment config - every retrieval call succeeds from the client's perspective and returns zero memories. The cognitive loop continues with an empty context.
Experiment 28 encountered this during the 100-turn session validation. A memoryIndex mismatch produced zero-result retrievals across the entire session. The session completed without errors; the results were simply wrong in a way that required inspecting the output to detect.
Mitigation: Monitor zero-result retrieval rates as a health signal. A healthy deployment should have near-zero zero-result retrievals for any query with non-trivial content. A spike in zero-result rate almost always indicates misconfiguration, index health degradation, or a corpus that has not yet been populated. This metric is cheap to instrument and catches the failure mode before it propagates.
knn shard constraint
What happens: Mixed-document shards cause ConjunctionDISI errors in OpenSearch knn queries.
Mechanism: OpenSearch 3.x knn vector search requires all documents in a shard to have the knn_vector field. If a shard contains documents without that field - for example, because a non-embedding document was written to a knn-enabled index - the knn query throws ConjunctionDISI errors that surface as retrieval failures.
Experiment 23 hit this when testing the abstraction engine's output. Abstraction documents written to the same index as embedding documents shared a shard, and the mixed shard broke knn queries across the index.
Mitigation: Use dedicated single-shard indexes for knn-enabled content. Do not write non-embedding documents to indexes that have knn mappings. The memory indexes in this system are already separated by content type (experience_events, memory_semantic, etc.); the failure mode arises when that separation is violated by a new document type. Index mapping validation at write time catches this before it affects query paths.
Abstraction symbolic label ceiling
What happens: All levels of the abstraction ladder produce the same dominant label token, making higher-level abstractions indistinguishable from lower-level ones by keyword alone.
Mechanism: The abstraction engine clusters memories at each level of its compression ladder using the same source corpus. At each level, the most frequent token in the cluster center becomes the label. When the corpus is dominated by documents that all share a common vocabulary (as operational telemetry documents tend to be), every abstraction level surfaces the same dominant token regardless of how the clusters have actually diverged semantically.
Experiment 22 measured this: all five ladder levels (experience, pattern, concept, principle, worldview) shared the same dominant token. The embedding-based cluster structure was meaningful - the centroids were distinct in vector space - but the keyword labels were not.
Mitigation: Embedding-based clustering rather than keyword-based labeling at higher abstraction levels. Experiment 23 tested this with nomic-embed-text embeddings: cluster centroids diverged visibly across levels, and label tokens shifted as the embedding space separated the abstractions. The keyword label ceiling is a labeling problem, not a clustering problem - the underlying abstractions are structurally sound even when the labels collide.
Operating within the envelope
Each failure mode above was discovered through systematic experiment, not from production incidents. That is the intended relationship between the experiment corpus and deployment: find the boundaries in a controlled environment, document them, and instrument for early detection in deployment.
The system behaves predictably within its designed operating envelope: a synthetic corpus with known importance distributions, knn-dedicated single-shard indexes, correct index configuration, and a scheduled re-consolidation job. Outside that envelope - different event distributions, mixed shards, misconfigured indexes, no re-consolidation - it degrades in the specific ways described here, not catastrophically and not silently (with the one exception of the silent misconfiguration mode, which requires active monitoring to detect).
The design philosophy article discusses the broader question of what the system claims and what it does not.