We subject a biologically inspired memory system for LLM agents (episodic traces, repetition strengthening, consolidation) to five pre-registered falsifiable gates with floor/ceiling arms and bootstrap CIs, all on consumer hardware with local models. Two popular mechanisms fail: repetition-weighted ranking loses to vanilla cosine retrieval (−0.44/−0.18, cue-specificity violation), and cosine similarity cannot separate same-fact paraphrases from same-template distractors (overlapping distributions), making embedding-only deduplication structurally unsound. Two mechanisms survive: selective consolidation compresses 350 exposures into 150 memories at statistical parity with store-everything; and a frozen 9B agent coupled to the memory converts API error messages into transferable competence, first-try success rises from 0.45 to 1.00 across 40 consecutive unseen episodes, while a memoryless control stays at 0.00 and an oracle reaches 0.95. Weights are never updated: memory is the only learning channel. Transfer replicates 3/3 seeds. We argue agent memory research needs control arms and pre-registration more than new architectures.
Matthias Raviotta (Sun,) studied this question.