Thinking of ACE? We Can Do It with Fewer Tokens

Two systems, ALTK-Evolve and ACE, improve LLM agent performance by creating reusable lessons from the agent's past actions, a process called agentic memory. Both methods agree on the importance of not over-compressing these learned lessons, instead counting their occurrences to retain detail. The primary difference is in delivery: ACE injects a complete playbook at every step, whereas ALTK-Evolve dynamically selects a smaller, relevant set of guidelines for the current task. This selective retrieval allows ALTK-Evolve to achieve similar or superior accuracy with a fraction of the token cost during inference.