Zhou et al. 2026 — localized memory update and search beat global reorganization on cost-utility, until long-context workloads flip the trade
From the paper's end-to-end operational-cost evaluation (RQ5, §4.5 — eight of its 12 memory systems, within a study spanning 5 benchmark workloads; promotion wording: "From an ablation across 12 memory architectures and 5 benchmark workloads" — corrected 2026-09-11: the ablations are the paper's §5, the cost rule is RQ5), the "Operational Scaling Rule": "Localized update and search yield the strongest cost–utility balance." Richer global-reorganization systems (named: Cognee, MemoryOS, Zep) pay off only when their upkeep avoids broad recomputation — "otherwise overhead offsets its gains" — and the trade reverses under long-context workloads, where "whole-memory coordination becomes the dominant cost driver."
Promotion wording (kept as history): "Latency detail carried honestly: figures relayed at capture level (LightMem ~3.7s vs Cognee ~116s / Zep ~155s per operation) came from an AI-summarized read of Figure 11, not extracted text — directional (order-of-magnitude gap favoring localized systems), not citable numbers. [unverified-quant]."
Latency detail, verified 2026-09-11 from the PDF's extracted text (§4.5, O7): "LightMem reaching 48.3 Normalized Utility at 3.67 s Avg. Operation Latency/Query" versus "Cognee and Zep exceed 84 utility only after 116.5 s and 155.1 s." The metric is Avg. Operation Latency/Query — memory construction time plus query time, amortized per query — so the figures are citable as that quantity, not as a per-operation latency. The order-of-magnitude gap the capture relayed is real; the paper's own reading of it is that "operational efficiency is governed less by whether a system uses structure than by how widely each write propagates through that structure."
Why the vault keeps this. It is general, external evidence for the vault's targeted-plus-periodic bet: forward-hooks + lint are "localized update"; a Dreams-style pass (claim-anthropic-dreams-nondestructive-reorg) is "whole-memory coordination." The first vault-lint's catch-count (31 hygiene / 0 truth-defects at n=63) points the same way — question-consolidation-pass-vs-revisit-protocol holds both pieces of evidence and stays open. The paper's own caveat transfers as a watch condition: the vault's coming embedding layer (question-embedding-layer-threshold-crossed) increases retrieval load per query, which is a step toward the workload regime where the paper says the trade begins to reverse. moc-peer-field-agent-memory.
Source
“Localized update and search yield the strongest cost–utility balance”