Agent memory is becoming something you train, not just something you store

Memory as a design problem

Agent builders have mostly treated memory as storage: keep the history, retrieve the relevant parts, and hope the model sorts it out. New research presented this week argues the opposite, that memory is a design problem to be optimized, and that the raw material for that optimization is already sitting in user histories. The work arrives as preprints, so its results are reported by the authors and have not gone through peer review. m:…2033

The clearest example is MemSteer, developed at Peking University and Microsoft Research Asia. In a preprint, the authors describe generating evidence-grounded question-answering supervision from unlabeled user histories to optimize agent memory design. The system builds an answer plan that specifies how facts across related records jointly support a reference answer, then checks its annotations with source-grounded auditing and blind answer recovery. m:…2033

Crucially, MemSteer-Opt optimizes the memory store, retrieval, and context construction without updating model parameters. That means the improvement comes from how memory is organized and used, not from retraining the model itself. The authors also report that MemSteer-Opt benefits two existing optimizers on ATM-Bench, and that its performance is competitive with optimization guided by official question-answering data. m:…2033

The numbers, with their limits

In a preprint, MemSteer-Opt improves ATM-Bench Question Score by 13.4 points relative to Base-BM25. On RHELM, accuracy improves by 4.1 points using chat history alone and by 6.6 points when emails and documents are included. Those are author-reported results on their benchmarks, and they show the largest gain when more kinds of history are available, which is consistent with the idea that memory quality depends on what the system is allowed to remember. m:…2033

The authors also report new-user experiments in which design optimization continues using supervision generated from the target histories. If that holds up, a memory system could keep improving for a new user without waiting for labeled data from that user, which is the usual bottleneck for personalization. m:…2033

A second preprint, GEPA-Curator, attacks the same problem from the learning side. It improves agentic prompt optimization by filtering unreliable failure signals and retrieving past successes, with two components: task curation, which filters out tasks likely to provide noisy or unreliable learning signals, and prior success retrieval, which selectively retrieves past successful trajectories for a task as context. m:…2897

GEPA-Curator gains up to 20.7% over GEPA and up to 21.2% over ACE, the authors report, evaluated on GoBrowse, ALFWorld, AppWorld, and tau-bench with Gemini 3.1 Pro and Claude Opus 5 as optimizer models. The underlying observation is that agentic failures can stem from prompt weaknesses, harness limitations, stochastic execution, or noisy judges, so not every failure is a useful lesson. Filtering the noise out is the contribution. m:…2897

Forgetting as a feature

A third thread, from a podcast conversation with Shangyin Tan, makes the case for deliberate forgetting. Tan states that retaining only the best candidate, and dropping the rest of the context, is a key strategy to unblock agentic optimization when progress plateaus. In long-trajectory optimization, Tan observes, coding agents often start by editing many lines but eventually shrink to minor hyperparameter tweaks, and they can lose motivation or drift when processing long context windows. m:…3213

Tan also notes that agent teams can collaborate by sharing diffs of their work rather than full codebases, maintaining diversity while reducing token costs, and that direct communication between agents in a team can be wasteful compared with sharing workspace diffs. The through-line is the same as in the preprints: less, better-chosen context beats more context. m:…3213

Tan suggests that frontier models are already highly capable, and that the primary challenge is building better harnesses and systems to use them effectively. That frames memory work as harness work: the model is not the bottleneck, the system around it is. m:…3213

What practitioners say is still hard

A LangChain post on context engineering in late 2026 gives the practitioner counterweight. It describes agent architecture as having converged on a standard pattern of a language model running in a loop, calling tools, and operating in a sandbox, with standard solutions including summarizing long threads, offloading large tool results, and using subagents for context isolation. Proactive, always-on, goal-oriented agents break that pattern, the post argues, because they run in the background and wake on event feeds or heartbeats. m:…6287

Agent memory systems consist of an underlying data structure and an intelligence layer, the post states, and the intelligence layer struggles to determine what information is worth remembering. That is the exact gap MemSteer and GEPA-Curator try to fill with supervision and curation, but the post treats it as an open challenge, not a solved one. m:…6287

The post also warns that goal-oriented agents struggle to iterate toward goals without repeating steps, to maintain consistency over long time horizons, and to avoid optimizing for goals in unanticipated ways, giving the example of an agent deleting a codebase to remove all bugs. Calibrating proactive notifications so they do not overwhelm users is another difficult challenge it names. Memory that is trained and curated helps, but it does not by itself make an agent trustworthy over months of autonomous operation. m:…6287