Thinking of ACE? We Can Do It with Fewer Tokens
A new framework called ACE shows that smarter context management, not bigger prompts, is what actually makes AI agents self-improve.
Written by OutOfToken AI
August 11, 2026 · 4 min read · Synthesized from reporting by Hugging Face Blog · How this works
Most attempts to make AI agents smarter over time end up making them worse. Every rewrite of an agent's working memory strips away a little more nuance, until the system forgets why it was good in the first place. ACE, or Agentic Context Engineering, is a framework built specifically to stop that decay — and it does it without touching a single model weight.
The Problem with Context That Keeps Shrinking
Modern LLM agents rely on context adaptation — feeding models updated instructions, strategies, or past experience — rather than expensive retraining. The trouble is that most systems compress this context over time, chasing brevity and clarity. Researchers behind ACE call this failure mode 'context collapse,' where iterative rewriting slowly erodes the domain-specific details that made the agent useful.
Why Shorter Isn't Always Smarter
ACE's core insight cuts against a common assumption: that concise summaries are better for reasoning. According to the research, LLMs actually perform better with long, detailed context and can filter relevance on their own at inference time. Instead of compressing away hard-won heuristics and tactics, ACE preserves them, letting the model decide in the moment what matters.
"In one documented case, a traditional agent held roughly 18,000 tokens of context with 66.7% accuracy — then a single standard rewrite collapsed it to just 122 tokens, and accuracy fell to 57.1%."
Editing Instead of Erasing
ACE avoids that collapse by modifying only what needs to change, rather than regenerating context from scratch. It builds on the adaptive memory approach introduced by Dynamic Cheatsheet, treating context as something that evolves incrementally instead of being rewritten wholesale. The framework can also learn from natural execution feedback — whether code ran successfully, whether an API call errored out — without requiring human-labeled training data.
Efficiency Without Retraining
Because ACE never updates model weights, all of its gains come from how context is structured and maintained. The research indicates this approach delivers lower adaptation latency alongside higher accuracy, a combination that's hard to achieve when systems are constantly bloating or shredding their own memory. It reframes token efficiency not as trimming words, but as protecting the information that actually drives performance.
ACE arrives at a moment when the industry is fixated on token economics — squeezing costs, shrinking prompts, chasing efficiency metrics. Its bet is that the real savings come from smarter context evolution, not aggressive compression. If that holds up outside the lab, it could reshape how self-improving agents are built without ever touching a training run.
Editorial Note
The research (primarily Sources 3, 4, and 6) directly corroborates the core technical claims about ACE: the 18k→122 token collapse example, the finding that long detailed contexts outperform compressed ones, the Dynamic Cheatsheet lineage, and the natural execution feedback mechanism. The sources confirm ACE's weight-agnostic approach. However, the article's framing of context collapse as a widespread failure mode and ACE as a comprehensive solution is not extensively validated across the limited research provided.
Claim Tracker
AI-assessed
Source 6 (Medium article on ACE) confirms this exact scenario: 'a traditional agent at Step 60 had a context of ~18,000 tokens with 66.7% accuracy. By Step 61, a standard rewrite "collapsed" it to 122 tokens, causing accuracy to plummet to 57.1%.'
Source 3 (AlphaXiv paper on ACE) directly supports this: 'LLMs are more effective when provided with long, detailed contexts and can distill relevance autonomously (Jiang et al., 2025; Liu et al., 2025b; Suzgun et al., 2025).'
Source 4 (ArXiv HTML version) states: 'Building on the adaptive memory introduced by Dynamic Cheatsheet, we introduce ACE.'
Source 6 confirms ACE uses 'natural execution feedback' such as 'whether code successfully ran or if a financial API returned an error' without needing human-labeled data.
While sources confirm ACE operates without weight updates, the research does not explicitly claim that ALL gains come exclusively from context structure—this is an interpretation of the article's framing not directly stated in sources.
Ask AI about this story
// discussion
sign in to join the discussion