Thinking of ACE? We Can Do It with Fewer Tokens

A New Approach to Agentic Memory

Researchers from IBM Research have detailed a new system for agentic memory, ALTK-Evolve, that achieves comparable or superior performance to the recently introduced ACE (Agentic Context Engineering) method while dramatically reducing inference costs. The work addresses a common failure point in multi-step AI agents: not a lack of knowledge, but an inability to reliably apply it. By learning from an agent's own history, both systems aim to improve task completion, but their differing approaches to delivering that learned experience have significant implications for token consumption, a key factor in the operational cost of deploying agents.

Both ALTK-Evolve and ACE build a repository of lessons from an agent's past actions without requiring weight updates or human labeling. Their fundamental disagreement is in memory delivery. ACE injects its entire, comprehensive 'playbook' of lessons into the context on every step an agent takes. In contrast, ALTK-Evolve treats delivery as a tunable dial, providing a small, fixed core of high-support guidelines and augmenting them with a few highly relevant lessons retrieved specifically for the task at hand. This shift from a constant, full-context injection to a selective, task-aware retrieval is the primary driver of its efficiency.

Performance and Cost Implications

The efficiency gains are substantial. On the AppWorld benchmark using a ReAct agent, IBM Research's internal testing showed ALTK-Evolve delivering superior accuracy on a strong model at roughly 40% of ACE's token cost. On a weaker model, it achieved parity on accuracy for about one-seventh the cost. This suggests that for less capable models, an unfiltered, comprehensive context can overwhelm rather than guide, particularly on difficult tasks where curated lesson retrieval proved more effective. The results indicate that how context is provided is as critical as what context is provided, offering a path to make complex agentic systems more economically viable for real-world deployment.

• Method Comparison: ACE injects a full playbook on every step; ALTK-Evolve uses selective, task-relevant retrieval.
• Cost on gpt-oss-120b: ALTK-Evolve used 116K tokens per task, while ACE used 777K for similar accuracy.
• Performance on DeepSeek-V3.2: ALTK-Evolve achieved 89.3 Task Goal Completion vs. ACE's 80.4, at less than half the token cost (263K vs. 634K).

The efficiency gains of ALTK-Evolve highlight a critical design principle for agentic AI: effective memory isn't just about storing experience, but about retrieving and delivering it intelligently. Injecting massive, unfiltered context can overwhelm less capable models and inflate costs, whereas calibrated, task-relevant retrieval can improve both performance and efficiency.