IBM Research’s ALTK-Evolve reduces agentic memory token usage by up to 86% with selective retrieval

IBM Research describes ALTK-Evolve, a system that builds agentic memory from an agent’s own trajectories and delivers it at inference without model updates or human labels. ALTK-Evolve and the previously published ACE both store lessons as itemized, non-summarized units with counters tracking how many independent episodes support each lesson.

The systems agree on core design principles: avoid compressing lessons into generic summaries, keep rich itemized records, and let the model distill relevance at read time. ACE organizes lessons into a single evolving playbook with per-bullet counters, while ALTK-Evolve consolidates lessons by clustering near-duplicates and merging clusters support-conservingly so the survivor inherits combined support counts.

ALTK-Evolve further extracts typed guidelines—strategy, recovery, and optimization—with causal attribution and provenance back to source trajectories, and stores them at subtask granularity to aid transfer across apps. Delivery is treated as a dial: a small fixed core of high-support guidelines is always available, and task-specific retrieval adds a handful of additional guidelines selected by cosine similarity, LLM guidance, or priority weighting. When a model has sufficient headroom, the full consolidated set can be injected.

On the AppWorld benchmark with the same base ReAct agent, IBM Research reports token usage and task goal completion for two models: DeepSeek-V3.2 and gpt-oss-120b. For DeepSeek-V3.2, ALTK-Evolve used 263K tokens per task while ACE used 634K, with ALTK-Evolve achieving equal or better accuracy across difficulty tiers. For gpt-oss-120b, ALTK-Evolve used 116K tokens per task while ACE used 777K, again matching or exceeding ACE’s accuracy, especially on hard tasks where curated retrieval outperformed a full playbook.

IBM Research notes that the stronger model benefits from the full consolidated set, whereas the weaker model is overwhelmed by large contexts, so selective retrieval is used to avoid distraction. They emphasize that the efficiency gains come from delivery calibration rather than lesson compression, and that the optimal amount to inject scales with model capability—a topic slated for a subsequent post.