Thinking of ACE? We Can Do It with Fewer Tokens

LLM agents often fail on realistic multi-step tasks not from lack of API knowledge but from not internalizing how to use APIs reliably—mis-paginating an endpoint, resolving the wrong person, or returning a value when none was asked for. ACE (Agentic Context Engineering) and ALTK-Evolve both address this as a form of agentic memory: turning an agent's past trajectories into reusable lessons and feeding them back at inference time, with no weight updates and no human labels.

Both systems refuse to compress lessons. ACE names failure modes like brevity bias and context collapse, and keeps a rich, itemized playbook with a helpful/harmful counter on every bullet, letting the model distill relevance at read time. ALTK-Evolve gives every distinct guideline a support count—how many independent episodes produced it—and never summarizes the store, treating a lesson discovered by five tasks as a different object from one that appeared once. Their core answer to whether to compress is the same: count them, don't collapse them.

The two differ in how memory is built and delivered. ACE grows one playbook through a Generator → Reflector → Curator loop with incremental delta updates and embedding-based deduplication. ALTK-Evolve clusters near-duplicate lessons and merges them support-conservingly so the survivor inherits the combined count, and it extracts typed guidelines (strategy, recovery, optimization) with causal attribution and provenance back to the source trajectory, at subtask granularity. This delivery difference is what shows up in the token bill.