Agent Memory Allocation: Tiers and Policies of Effective Agent Memory

Agent memory is theorized and measured almost entirely at the context window. Yet a deployed agent's effective memory spans several tiers — a boot-resident image (contract files, skill and tool catalogs, memory indexes), pointer- resident and paged-searchable filesystem content, and tool-gated stores — whose relative utility has, to our knowledge, no published cross-tier measurement. We make three contributions. First, we formalize agent memory as a set of tiers characterized by capacity, access cost, persistence, addressability, and staleness risk, governed by five policies: residency, paging, promotion, eviction, validation. The policy names are borrowed from classical cache and virtual-memory design (Smith 1982; Denning 1968, 1970); what the tier formalization adds beyond that vocabulary is the content-versus-pointer residency distinction and the staleness-risk property, which has no cache analogue because cache paging is faithful and digest paging is lossy. Second, we show the vocabulary expresses four existing systems (MemGPT, Pichay, the Claude Code harness, and a production workspace) without a residual category, and that the validation cell is empty for the two published systems we read. That is a property of two selected neighbours, not of the literature: MemoryBank's Ebbinghaus-inspired decay mechanism fills the same cell. Third, we run a pre-registered falsification program over one single-operator agent ecosystem, with predictions frozen and analyzers committed before any data were read. Its results, each adjudicated mechanically against criteria fixed in advance: promotion of a fact to the durable memory directory does not produce later returns (2 of 157 eligible promoted files ever re-read — refuted); associative search-then-read beats curated layered disclosure on accuracy (72.5% vs 47.1%, +25.5 pp), on error rate (27.5% vs 52.9%), and on marginal token cost, while costing 1.41× more on total tokens, the mechanism being that the authored index routed the policy to the correct file on 52.0% of questions against grep's 80.4% (supported on this corpus). That finding has since been replicated in public at power: a pre-registered replication over 141 releasable documents and 120 frozen questions returns search +12.5 pp on accuracy (63.3% vs 50.8%), a wrong-stop tax of 34.2% against 18.3% under a symmetric rule the original lacked, and a localization advantage of 75.0% against 62.5% — all three predictions passing, with the corpus, questions, answers, harness and adjudicator shipped for re-running. Its overall verdict is revised, on an orthogonal fifth prediction about token headroom that failed on the harder of two measures (§6.5); the replication supersedes the underpowered demonstrator, which adjudicates nothing (§7 Limits (b′)). And the frame's own residency-depth ordering is inverted in recorded-but-under-powered estimates, where searchable content out-returns pointer-resident content, although a post-hoc diagnostic shows that inversion is substantially carried by reads performed by the sessions that were editing the files. We report the negative, self-inverting, and self-correcting results as first-class, and argue that the method, a rigorous instantiation of pre-registration and mechanical adjudication in a systems setting, is itself a contribution.

Keywords: context engineering; agent memory; retrieval; progressive disclosure; pre-registration; LLM agents.

Discipline carried from the governing program and observed throughout: every quantity cites a source path; no outcome of unadjudicated work is anticipated (the covered C26/C27 re-runs, the registered-but-unadjudicated successor conjectures c26b and c28b, and the candidate conjecture c30 are pre-judged nowhere); no self-assigned quality score appears.