Agent memory isn't a switch, it's a dosage calibrated to the model

The memory doesn't turn on; it's calibrated. Researchers tested eight models, from a compressed 30-billion-parameter model to frontier proprietary systems, and found that agentic memory does not uniformly improve performance. The method, called ALTK-Evolve, lets an agent learn from its own past trajectories: it distills guidelines for reuse and injects them at inference time, without weight updates and without human labeling. The main conclusion: the right dosage depends on the model's capability tier, not on parameter count alone.

The experiments revealed three recurring patterns. Strong models with headroom want the full guideline set, including rare edge-case lessons, because they have the capacity to absorb and apply all of them; DeepSeek-V3.2 (671B MoE) climbed 9.5 percentage points on task completion when given the full set. Smaller or weaker models drown in a large set; for them, a tight, safe core plus targeted retrieval for the task works best. gpt-oss-120b (117B MoE) gained 16.1 percentage points with the selective approach, while the full set yielded less improvement and cost roughly 50% more tokens. Saturated models show no measurable improvement; GLM-5 (745B MoE) fell into this category. The researchers stress that the classification describes an observation, not a proven cause — the model may already be near its ceiling, the guidelines may not have hit the remaining failure modes, or the model may not have applied the guidance effectively.

"Memory" here is not a replay of a past transcript, but a set of guidelines — successful strategies, mistakes to avoid, edge cases — distilled from the agent's own trajectories. The loop is simple: the agent attempts tasks and produces trajectories; ALTK-Evolve extracts guidelines from both successful and failed runs; unifies them into a reusable set; and at inference the agent receives either the full set or a task-relevant selection. Because there are no weight updates, learning changes only the guidelines available to the agent, making adoption cheap and portable across the models tested.

The evaluation ran on AppWorld, 585 multi-step tasks (168 test_normal and 417 test_challenge) across 9 simulated applications — calendars, messaging, payments, and so on. Two metrics were measured: TGC, whether the agent completes each task in full, and SGC, whether every variation of a scenario passes, a stricter all-or-nothing bar. Both memory configurations draw from the same guideline set mined once from the AppWorld training split alone; only the injection method at inference changes. Curated retrieval proved both the most accurate and the cheapest: gpt-oss-120b achieved its 16.1-percentage-point jump with only a 5% token overhead, and prompt caching keeps even the full set viable in production.