Agent memory improved most on temporal and multi-hop questions
The biggest measured gains are +29.6 points on temporal queries and +23.1 on multi-hop reasoning. Those two categories are precisely what breaks when an agent has to remember a real user.
A progress report on agent memory puts the largest improvements at +29.6 points on temporal queries and +23.1 on multi-hop reasoning.
Those two categories are not arbitrary. Temporal means when — what did the user say last week, what has changed since, what is now stale. Multi-hop means chaining two remembered facts to answer a third question. Both are trivial for a person with a notebook and both have been badly handled by systems that treat memory as a similarity search over past text.
The reason they break together is that vector retrieval has no native notion of time or of composition. Nearest-neighbour search over embeddings returns things that are similar, and "most recent" and "related to the thing I just retrieved" are not similarity relations. Gains on both at once suggest memory systems are acquiring structure rather than just more storage.
The caveat is what a point means. These are benchmark deltas, and agent memory benchmarks are young enough that the tasks are still being negotiated alongside the results. A 29-point gain on a metric the field defined last year is progress; it is not the same as a user noticing.
Still, it is the right pair of things to be fixing. Memory is what makes an agent's second conversation better than its first, and memory provenance is now on the list an enterprise buyer will ask about.
Mem0 — State of AI Agent Memory 2026: Progress Benchmark Report → · arXiv — From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review →