AI Signal 508
Thinking of ACE? We Can Do It with Fewer Tokens
↗
ALTK-Evolve and ACE both store agentic lessons without compression, but ALTK-Evolve trims the inference context to fit model capacity, cutting token usage dramatically.
Engineers running LLM agents can lower inference costs by up to 60 % on strong models and 85 % on weaker ones while keeping accuracy essentially unchanged. The approach requires adding a retrieval-selection layer, but it avoids the overhead of feeding an ever-growing full playbook at every step.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Both systems keep each learned lesson as a distinct, count-tracked guideline rather than summarising them.
ACE always injects the entire evolving playbook each inference step; ALTK-Evolve sends only a core plus a task-specific subset.
In benchmark runs ALTK-Evolve used 263 K vs 634 K tokens for DeepSeek-V3.2 and 116 K vs 777 K tokens for gpt-oss-120b, with comparable accuracy.
THE READ
What the cluster adds up to.
Agentic memory techniques let LLM agents refine their behavior from prior runs, and both ACE and ALTK-Evolve treat each discovered guideline as a separate, count-annotated entry. This design choice avoids the “compression” pitfalls of collapsing many experiences into a few vague rules, preserving the richness of the learned material for later reuse. The shared premise is that a model should decide relevance at read time rather than rely on a pre-summarised context.
The two systems diverge in how they construct the lesson store. ACE builds a single, evolving playbook through a Generator-Reflector-Curator loop, clustering near-duplicate lessons and merging them while preserving support counts. ALTK-Evolve, by contrast, aggregates lessons with support counts but does not describe a de-duplication step, instead focusing on keeping each guideline individually retrievable for later selection.
Delivery is where the token disparity originates. ACE injects the full playbook into the model’s context on every reasoning step, regardless of model size or task complexity. ALTK-Evolve treats the context as a configurable dial, delivering a compact core of high-support guidelines and augmenting it with a handful of task-relevant entries chosen by similarity or LLM-guided weighting, allowing the model to operate within its token budget.
Empirical results on the AppWorld benchmark illustrate the impact: for the stronger DeepSeek-V3.2 model, ALTK-Evolve required 263 K tokens per task versus ACE’s 634 K, and for the weaker gpt-oss-120b model the figures were 116 K versus 777 K. Accuracy stayed within the noise margin, with ALTK-Evolve matching or slightly edging ACE, showing that the token savings do not sacrifice performance.
For engineers, adopting ALTK-Evolve means adding a retrieval and relevance-ranking component to the inference pipeline, which incurs development effort but yields substantial cost reductions on token-priced services. The approach may falter if the selection mechanism omits a guideline essential for a particular subtask, leading to failures that ACE would have avoided by always providing the full playbook. Monitoring retrieval effectiveness and fallback strategies will be key to maintaining reliability.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER