Beyond Downstream QA: Rethinking How We Evaluate Agent Memory
Sebastian-Antonio Toma and colleagues (Amazon Web Services and Technical University of Cluj-Napoca) show that measuring agent memory through downstream QA accuracy hides most memory errors, and release FEB-30 to inspect the memory store directly.
Ask this paper
Why QA accuracy misleads. It mixes memory correctness with retrieval and generation, so a correct answer can hide a broken memory and a wrong answer can blame a correct one.
Benchmark construction. 30 fact-evolution trajectories across 6 diagnostic categories and 3 difficulty levels. Documents, emails and conversations are generated from known fact trajectories, which gives deterministic ground truth without human annotation.
Divergence. On a representative memory system, store-level and QA-level scores differ by 16 points on average, and 44% of trajectories exceed the divergence threshold.
Hidden failure. The system labels memory operations (create, update, reinforce, supersede) correctly only 20% of the time, and QA-based evaluation does not detect this at all.
Abstract
Agent memory benchmarks evaluate whether an agent's responses reflect correct knowledge, measuring downstream question-answering accuracy as a proxy for memory quality. We show this evaluation paradigm is incomplete: it conflates memory correctness with retrieval and generation quality, producing both false positives where correct responses mask broken memory and false negatives where incorrect responses blame correct memory for retrieval failures. We introduce FEB-30, a benchmark of 30 fact-evolution trajectories spanning 6 diagnostic categories and 3 difficulty levels, constructed via paradigm inversion: documents are generated from known fact trajectories, yielding deterministic ground truth without human annotation. Each trajectory defines how a fact should evolve across multi-author artifacts (documents, emails, conversations), enabling direct inspection of the memory store after each ingestion step. Applying FEB-30 to a representative memory system, we demonstrate that intrinsic and extrinsic evaluation diverge by an average of 16 percentage points, with 44% of trajectories exceeding meaningful divergence thresholds. Diagnostic decomposition reveals that the system correctly classifies memory operations (create, update, reinforce, supersede) only 20% of the time, a failure entirely invisible to extrinsic evaluation, and that failure modes differ systematically by category and difficulty. We release the benchmark, document generation pipeline, and scoring framework to support reproducible evaluation of agent memory systems.