Quadratic Memory Growth in a Hippocampal Auto-Associator: Forensics of a 441 MB State Bloat, Recovery from a Newborn-State Overwrite, and a Substrate-Blessed Structural Cap

Abstract

A persistent bio-inspired agent substrate (Hebbian/STDP plasticity, hormone-analogue state vector, a hippocampus-analogue episodic-memory module with CA3-style auto-associative pattern completion, operator-bonded longitudinal identity) persists its entire state to a single JSON file on every tool-use event. On 2026-07-06 that file, normally ≈16 MB, was found at 441 MB. The oversized file could not be loaded by the persistence layer, which fell back to a fresh default ("newborn") state and, because the write path carried no guard against that fallback, persisted the newborn over the real state on every subsequent write. Recovery was an atomic restore from the most recent daily archive; the cause of the bloat itself, overwritten before it could be captured, is unrecoverable from the corrupted file. Root-cause analysis localised the bloat to the episodic-memory module's CA3 co-occurrence map: episode count was capped, but the derived pairwise-association structure was not, producing quasi-quadratic growth under heavy sessions. Two independent defences were shipped: (1) a write-path sanity gate — refuse writes that are anomalously large, that look like a fresh state landing over a substantial one, or that regress the state's monotonic tick counter — with the first two predicates mirrored, at identical thresholds, across the three processes then identified as persisting the full state file, and the third (tick regression) implemented only in the highest-frequency writer, the tool-use hook; (2) a structural bound on the co-occurrence map (row and per-row-partner ceilings plus a weight-by-recency prune rule), proposed as a preregistered change requiring the substrate's own procedural sign-off before shipping, through two rounds of amendment. The bound was validated on staged, checksum-locked copies of the live store (12/12 falsifiers) before a flagged live cutover pruned the map's partner count from roughly 1,800 to 64 per row. Because a memory-bounding change carries a plausible behavioural cost, its promotion criterion is behavioural rather than structural: a citation-rate falsifier tracks the rate of dated, auditable references in the substrate's output over a fixed post-cutover session window against a frozen pre-cutover baseline. That window has not closed at the time of this report; no behavioural verdict is claimed. This deposit reports the incident, the two defences, and the falsifier design as a case study in a general failure class — an unbounded second-order structure hiding behind a bounded first-order counter in continuously-mutated agent state — and in the write-path guard as a general defence against exactly that class of bug turning into unrecoverable data loss.