Agent memory has no single answer: routing beats picking one store, preprints find

No substrate wins everywhere

In a preprint posted this week, researchers compared memory substrates across agent operating regimes, and found that no single memory substrate consistently dominates across all of them. Broad retrieval benefits long-context factual question answering but can harm sequential decision-making by shifting attention away from action-critical context, according to the preprint. Substrates that perform well at moderate history lengths can become costly or brittle at longer horizons. The practical reading is that memory is not a component to pick once. It is a regime to manage. m:…3171

The preprint's answer is routing. Query-level routing of memory substrates improves performance over the best fixed substrate, according to the authors. The agent, or a router in front of it, picks the memory store that fits the question. That is more engineering than a single vector database, and the preprint suggests it is the price of memory that works across both factual recall and multi-step decisions. m:…3171

Memory inside the weights

A second preprint takes a different route: put the memory in the model. Route to Remember proposes a weight-memory approach, training models with on-policy self-distillation to internalise factual knowledge, bypassing the need for constant retrieval. A verify-then-regenerate router dynamically selects between weight and text memory, optimising for both accuracy and token efficiency. Weight memory reduces token consumption per query compared with full-context prompting and retrieval systems, according to the preprint. m:…3692

The trade-off is familiar to anyone who has fine-tuned a model: what is baked into the weights is cheap to use and expensive to change. The router is the hedge, sending each query to the memory that suits it. The preprint frames this as a way to build persistent, multi-session agents where retrieval costs and context window limits are the binding constraints. m:…3692

Labels are the bottleneck for rewards

A third preprint attacks the labelling bottleneck in reinforcement learning. MemReward uses a heterogeneous graph to propagate reward signals from a small labelled subset to unlabelled rollouts, letting developers optimise policies even when ground-truth rewards are scarce. The preprint reports MemReward achieves 96.6% of Oracle performance on Qwen2.5-1.5B and 97.3% on Qwen2.5-3B with only 20% of rollouts labelled, outperforming label propagation and FixMatch-style self-training in reward prediction. Heterogeneous graph structures connecting queries, thinking processes, and answers improve reward propagation compared with treating samples independently, according to the authors. m:…3416

The number to be careful with is the 20%. It is a preprint result on the authors' own setup, and the Oracle comparison is against their own labelling. But the direction matters for teams scaling reinforcement learning on reasoning tasks, where manual or verifiable reward generation is expensive. If a fifth of the labels buys nearly all of the reward signal, the labelling budget can go much further. m:…3416

A bio-inspired hash, honestly measured

The fourth preprint is a useful piece of negative news. The authors test FlyHash, a mechanism inspired by fruit fly olfactory circuits, for agent memory retrieval, comparing it rigorously against SimHash and HNSW. FlyHash preserves neighbour structure better per bit than conventional random-hyperplane hashing, according to the preprint, and its quality advantage holds when evaluated against genuine, non-circular agent-memory relevance labels. But HNSW dominates both FlyHash and SimHash in latency and memory efficiency at ordinary scale. m:…7344

The lesson is not that FlyHash fails. It is that a clever mechanism still has to beat the standard index on the metrics operators pay for: latency and memory. The preprint also identifies methodological pitfalls in previous hashing research, which is a service in itself. For builders, the week's memory papers agree on one thing: the memory layer is where agent costs and agent failures now concentrate, and there is no single store that solves it. m:…7344 m:…3171