Metabolic Memory Bench (MMB-v0.3): Beyond Query-Driven Retrieval — Measuring Endogenous Initiative, Functional Forgetting and Relational Continuity in a Living Agent Log

Benchmarks for agent memory — LoCoMo, LongMemEval, BEAM — share what we call the Librarian Fallacy: they evaluate memory only as a reactive, query-driven store (prompt → search → extract → respond). They leave unmeasured the operations that define long-horizon companionship: initiating contact without a trigger, letting noise decay while invariants consolidate, surfacing the past unprompted, recognising the counterpart before parsing words, and internalising their rhythm. We introduce Metabolic Memory Bench, an architecture-agnostic protocol computed from four append-only telemetry streams whose schema is published in full, and report the first reference run: 25 days, 4,678 turns and 713 logged initiative decisions from a homeostatic agent in continuous use with two human actors. The agent spoke first 403 times; 86% of those messages were answered within an hour, 45% within five minutes, with 0.7% of speech falling into the actor's closed presence window. Replaying the log under three regimes — live, live followed by 14 silent nights, and without trace confirmation — shows that forgetting in this class is governed by confirmation, not by time: constant contact retained 78% of episodes verbatim; two weeks of silence consolidated rather than erased (130 traces consolidated, 64 dissolved); removing confirmation retained 23% and engaged gist distillation. Two metrics returned negative results and one was found tautological on first contact with data; all are reported. Axes are never collapsed into a single score. Memory is not what you can retrieve. It is what you have become. Supplementary: aggregate results for all three regimes (mmb_results_*.json). Harness and reference system are proprietary, available under NDA on request.