IBM Research published ALTK-Evolve, an agentic memory system that consolidates an agent's learned failure modes and lessons into selectively retrievable guidelines—achieving up to 7x lower token consumption than ACE (Agentic Context Engineering) on multi-step reasoning tasks. On a weak model (gpt-oss-120b), ALTK-Evolve matched ACE's 54.8% goal completion rate on AppWorld benchmark while consuming 116K tokens per task versus ACE's 777K. On stronger models (DeepSeek-V3.2), ALTK-Evolve improved accuracy (89.3% vs 80.4%) while cutting costs to 40% of ACE's 634K tokens.
Both systems learn from agent trajectories without weight updates or human labels. ACE and ALTK-Evolve agree on a core principle: never compress an agent's hard-won lessons into brief summaries; instead, preserve rich, itemized stores with support counts showing how many independent episodes produced each guideline. The divergence is in delivery: ACE injects the entire comprehensive playbook on every inference step; ALTK-Evolve uses selective retrieval, sending a fixed core of high-support guidelines plus task-specific lessons selected via cosine similarity or LLM ranking.
Token efficiency comes from this retrieval-vs.-injection trade-off. On harder tasks where models must pick the right lesson rather than wade through generic instructions, curated selection pulls ahead. ALTK-Evolve also extracts typed guidelines (strategy, recovery, optimization) with causal attribution back to source trajectories and at subtask granularity, enabling transfer across agent domains. The work tests both systems on AppWorld, a multi-step agent benchmark spanning bill-splitting, song search, and order reconciliation across simulated APIs.
For architects: agentic memory cost matters at scale. If you run dozens of concurrent agents over months, token efficiency per task determines capex. ALTK-Evolve's selective retrieval strategy transfers across weaker and frontier models. The typing and causal provenance enable reuse across unrelated agent domains—a lesson learned by one app can serve another. Watch for production deployments; at 7x token reduction on weak models, the economics shift who can afford persistent agentic memory.