Agent memory evaluations often miss key failures in knowledge handling
Beyond Downstream QA: Rethinking How We Evaluate Agent Memory
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Many computer systems that remember facts are tested by checking if they answer questions right, but this can hide memory problems or blame memory for other errors. The authors created a new test called FEB-30 that uses carefully made documents with known fact changes to check memory more directly and clearly. Using FEB-30, they found that existing tests often give misleading scores and miss important mistakes in how systems update facts over time. They share the test and tools to help others accurately evaluate memory in smart agents.
What this means in practice
- •For conversational ai developers: Test agent memory more accurately using FEB-30 to identify hidden errors in how knowledge updates over time.
- •For knowledge management engineers: Directly inspect memory states after fact changes with FEB-30 to improve systems that track evolving information.
Authors
Sebastian-Antonio Toma, Jean de Kernier, Mike Ply, Oliviu Matei
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.