.png)
IBM Research released ALTK-Evolve, a framework for turning an agent’s past trajectories into reusable guidance, and paired it with an evaluation showing that memory is not a universal win. The central result is practical: the amount and shape of agentic memory should be tuned to the model, not bolted on uniformly.
ALTK-Evolve is an on-the-job learning loop for agents. It distills lessons from completed trajectories into guidelines, then injects those guidelines back into future prompts at inference time. That keeps the system in the prompt/memory layer: no fine-tuning, no weight updates, and no human annotation pass.
That design matters because it makes the memory mechanism operationally lightweight. The agent can improve from its own mistakes without retraining, but the improvement path is constrained by what the base model can actually use. The paper’s evaluation across eight models makes that constraint explicit.
The most useful result in the release is the nonlinearity across model tiers. The same memory strategy does not produce the same behavior everywhere.
Strong models with headroom benefit from the full guideline set, including rare edge-case lessons. DeepSeek-V3.2, a 671B MoE model, improved by 9.5 percentage points in task completion when given its full self-mined guideline set. In that regime, the model has enough capacity to absorb a large memory payload without getting overloaded.
Smaller or weaker models behave differently. For them, a long guideline list becomes noise unless it is aggressively filtered. A compact core of high-confidence guidelines plus task-relevant retrieval performs better than dumping the whole memory store into context. gpt-oss-120b, a 117B MoE model, gained 16.1 points with selective retrieval, while the full guideline set cost roughly 50% more tokens and delivered less benefit.
Already-saturated models show no measurable gain. That is the most important caution for production teams: if the base model already solves the task well enough, adding memory may just add latency and cost without moving quality.
The evaluation makes a clear case for curated retrieval as the default memory policy for mid-tier systems. It was both the strongest accuracy/cost tradeoff and the most production-friendly option in the reported results.
The notable detail is token efficiency. On gpt-oss-120b, the selective approach produced the biggest improvement while adding only about 5% tokens. That makes it materially different from a naive “append all guidelines” implementation, especially once you factor in prompt caching. With caching, even the full guideline set becomes affordable in some deployments, but the raw token economics still favor retrieval when the model does not need the entire memory dump.
That suggests a simple deployment heuristic: use full guideline injection only when the model is demonstrably able to exploit it, and otherwise treat memory like any other retrieval system — rank, filter, and cap aggressively.
ALTK-Evolve is attractive because it turns agent experience into a first-class artifact, but the evaluation also shows why this should be treated as a model-specific control surface rather than a blanket feature. The right question is not whether an agent should have memory; it is how much memory a particular model can use before the marginal guideline turns into context clutter.
For production teams, the useful workflow is to benchmark three configurations on representative tasks: full guideline injection, compact core plus retrieved task-specific guidance, and no memory at all. If the model is in the saturated regime, ship nothing. If it has headroom, full memory may be justified. If it sits in the middle, curated retrieval is the likely default.
That framing makes ALTK-Evolve more than a self-improvement wrapper. It is a reminder that agent memory is another capacity-sensitive system component, and the optimal setting depends as much on the model underneath as on the memory store above it.
