The agent memory numbers published over the past week are high enough to be uninformative. Mem0's benchmark breakdown, posted about five days ago, reports LoCoMo at 92.5% and LongMemEval at 94.4% — and BEAM 1M at 62%. EverMind's August update claims state-of-the-art across four long-term memory evaluations, including 90.04% recall on HaluMem, alongside LoCoMo, LongMemEval and PersonaMem v2.
Two benchmarks sitting in the low-to-mid 90s across competing systems have stopped discriminating between them. The 30-point drop on BEAM is the number that still carries information.
The gains are on the write path, not the retriever
The more useful detail in Mem0's state-of-the-field post is where its April 2026 algorithm actually improved. Built on single-pass hierarchical extraction and multi-signal retrieval, its two largest gains over the prior algorithm were temporal queries (+29.6 points) and multi-hop reasoning (+23.1). Those are the categories where facts accumulate, change, and reference each other — exactly the cases flat top-k retrieval handles badly.
The design change behind that is a write-path decision: agent-generated facts are now stored as first-class, so the model's own confirmations and recommendations carry the same weight as user-stated facts. If your memory layer only ingests user turns, half the conversational state never gets written. No amount of retrieval tuning recovers it.
Mem0 also reports roughly 6,956 tokens per retrieval call on LoCoMo, and flags explicitly that this isn't directly comparable to the ~26,000 tokens per conversation in its 2025 paper — different units. Worth noting because per-call token cost is the metric that governs whether a memory layer survives contact with production traffic, and almost nobody reports it next to accuracy.
The replacement benchmarks measure memory as an action
Three efforts converge on the same critique. Microsoft's STATE-Bench is open source and memory-agnostic, with 450 tasks across customer support, travel and shopping covering policy compliance, information synthesis and multi-step reasoning; it asks whether agents improve with experience rather than whether they can recall a fact. MemoryArena frames the problem as multi-session Memory-Agent-Environment loops, arguing that existing evaluations assess memorization and action in isolation when the two are tightly coupled in practice. AMA-Bench makes a parallel point from the other direction: current benchmarks are dialogue-centric, while real agent memory is a continuous stream of machine-generated agent-environment interactions. LoCoMo-Plus and LongMemEval-V2 extend the originals, though Mem0 notes neither has displaced them.
A recent survey (arXiv 2603.07670) formalizes the shape this implies — memory as a write–manage–read loop coupled to perception and action — and points to work treating store, retrieve, update, summarize and discard as callable tools optimized end-to-end with RL.
Who this matters to
If you are choosing a memory layer this quarter, a LoCoMo number tells you almost nothing now. Ask for BEAM, a memory-agnostic harness like STATE-Bench, and tokens per retrieval call. If you are building one, the leverage is in what you decide to write and when you invalidate it — not in your reranker.