Freezing Opportunity: Matched-Unit Attribution for Shared-Budget Agent Memory Control

An agent-memory policy allocates a shared budget across retrieval, verification, summarization, and writing, so each decision changes which information actions remain affordable later. Policy comparisons become confounded when prompts, tools, candidate pools, admissible actions, or accounting rules also change. We introduce opportunity-matched attribution, a protocol that fixes this interface contract and matched evaluation units, then resolves a policy gain through four pre-assigned contrasts: exact-parity deployment, paired same-information control, shared-training-trace temporal control, and randomized feasible-action intervention. The resulting estimands distinguish gains due to controller organization from gains due to information access, temporal representation, or score ordering. LedgerGate instantiates the protocol as a support-aware five-action scheduler whose role-typed state limits aliasing, reliability particles update the value of expensive information, reserve-aware scoring allocates one token/time ledger, and deterministic fallback governs weak-support states. On LOCOMO v2, ToolBench-POMDP, and SWE-verify, LedgerGate improves tight-budget success over a pre-specified MLP by \(+2.2/+3.7/+2.4\) percentage points. At the medium budget, it remains within 0.5 points of Always-retrieve while reducing memory-action tokens by a 31.7% cross-benchmark median. Across \(20\) matched seeds, LedgerGate retains \(+1.6/+1.8/+1.2\) points over the same-information controller and \(+0.9/+1.1/+0.8\) over a parameter-matched GRU, with positive seed-first 95% intervals on every benchmark. Feasible-state interventions reach 82–87% sign-correctness for RETRIEVE and 74–79% for VERIFY, the actions carrying most resource pressure. Opportunity matching thereby converts bundled agent-system comparisons into controlled policy attribution and makes competing explanations empirically separable.