IBM Research reports ACE-style agent memory can be delivered at a fraction of the token cost

IBM Research has published a comparison of two agentic-memory systems that let a language-model agent learn from its own past trajectories without any weight updates — its own ALTK-Evolve, and the Agentic Context Engineering (ACE) system published earlier. Both turn an agent's past attempts into reusable lessons and feed them back at inference time; the difference the authors highlight is how the lessons are delivered, and what that means for the token bill.

The two systems agree that lessons should not be compressed into a tidy summary, IBM said. ACE calls this trap 'brevity bias' and 'context collapse' and keeps a rich, itemised playbook, with a helpful-or-harmful counter on every bullet. ALTK-Evolve keeps a support count on every guideline recording how many independent episodes produced it, so a lesson discovered by five tasks is a different object from one that appeared once.

The two differ on what reaches the model at inference time. ACE injects the comprehensive playbook on every step regardless of model or task. ALTK-Evolve treats delivery as a dial: a small fixed core of high-support guidelines is sent every time, extended per task with a handful selected by cosine similarity or an LLM, or by the full consolidated set when the model has the headroom to use it.

On the AppWorld benchmark, IBM ran both systems on the same base ReAct agent. With DeepSeek-V3.2 as the backbone, ALTK-Evolve reached 89.3 task-goal completion and 80.4 subtask completion against ACE's 80.4 and 73.2, while using 263,000 tokens per task versus ACE's 634,000 — about 40 percent of the cost. With gpt-oss-120b, the two systems' accuracy was effectively tied at around 55 to 56 task-goal completion, but ALTK-Evolve used about 116,000 tokens per task versus ACE's 777,000.

Broken down by difficulty, IBM said, ACE's full playbook has the edge on the easy and medium AppWorld tasks with gpt-oss-120b, but retrieving the right lesson pulls ahead on the hard tasks where the aggregate is decided; on the stronger DeepSeek-V3.2 backbone, ALTK-Evolve leads on easy, hard and overall.