Thinking of ACE? We Can Do It with Fewer Tokens

The blog introduces ALTK-Evolve and compares it to ACE (Agentic Context Engineering), both of which are forms of agentic memory that convert an agent's past trajectories into reusable lessons, injected at inference time without updating weights. The two systems agree on the core principle of not compressing lessons: ACE uses a comprehensive, evolving playbook with per-bullet helpful/harmful counters, while ALTK-Evolve consolidates similar lessons into clusters and maintains a support count for each guideline, never summarizing the store down. They differ in how memory is built and delivered: ACE grows a single playbook through a Generator, Reflector, and Curator loop with delta updates and embedding-based deduplication, whereas ALTK-Evolve clusters near-duplicates, supports merging with support-conservation, and extracts typed guidelines (strategy, recovery, optimization) with provenance at subtask granularity. The key difference is delivery: ACE injects the full playbook on every step, while ALTK-Evolve sends a small fixed core of high-support guidelines plus a per-task selection, or the full set if the model can handle it. On the AppWorld benchmark with the same ReAct agent, ALTK-Evolve achieves higher or comparable TGC and SGC scores with far fewer tokens: for DeepSeek-V3.2, TGC 89.3 vs 80.4 and 263K vs 634K tokens per task; for gpt-oss-120b, TGC 56.0 vs 54.8 and 116K vs 777K tokens. The accuracy gains come from retrieving a few guidelines per task instead of injecting the whole playbook, and the by-difficulty analysis shows that on hard tasks, curated retrieval outperforms the full playbook, while on easier tasks, the full playbook may help more but the difference is small. The blog notes that ACE's own efficiency is in building context cheaply, while ALTK-Evolve's is in serving it, and that the calibration of delivery is what drives the token savings. The ALTK-Evolve library, including extraction, consolidation, and retrieval pipeline, is available, along with a full technical report.