Skip to content

Blog

Agents fail from amnesia: context as a living playbook

Agent failures are context failures. The ACE framework treats context as a playbook that grows through small updates instead of rewrites, and small open models gain the most from it.

Written by

Context engineering

Published: 25 November 2025 Updated: 2 min read

Your AI agents aren’t failing because of the model. They’re failing because of amnesia.

Most teams solve the wrong problem: optimizing prompts instead of architecting memory.

Every agent failure is a context failure

Every AI agent failure is a context failure, not a model failure. We treat context as disposable and rewrite it from scratch. This creates two problems.

Brevity bias. We compress strategies into terse summaries, losing the nuanced operational evidence that actually makes agents reliable.

Context collapse. Each rewrite erases lessons learned. Agents reset. They don’t improve. They repeat the same mistakes, like a developer with amnesia after every deploy.

ACE treats context as a living playbook

ACE (Agentic Context Engineering) solves this by treating context as a living playbook, not a static prompt. It’s built on three roles working in concert:

  • Generator: executes tasks and produces reasoning traces and tool-call logs.
  • Reflector: distills actionable lessons from execution traces. It identifies what works and what fails, and builds the knowledge base.
  • Curator: integrates lessons as tiny, modular “delta updates”, not wholesale rewrites. It keeps a persistent, deduplicated playbook that grows smarter with every execution.

Instead of context collapse, you get context progression.

What ACE measured

On AppWorld, ACE matches top-ranked agents on average, but beats them on harder test splits.

  • +10.6% performance gain on agentic tasks vs. strong baselines
  • +8.6% accuracy improvement on finance benchmarks
  • 86.9% lower adaptation latency on average
  • Over 75% fewer rollouts

And it does this with smaller, open-source models, not proprietary LLMs.

The advantage is the context architecture

For years, we chased bigger models. ACE flips the equation. It shows that the competitive advantage isn’t the model weight. It’s the context architecture.

Small, open models with a well-engineered context system now approach or match commercial agent stacks. That changes accessibility, cost, and control.

If your team treats prompts as one-off config, you’re losing to teams that treat context as a system.

From rewrites to compounding updates

The old way: rewrite context, lose history, start fresh, reset.

With ACE: modular updates, preserved evidence, compounding learning.

This is what context engineering actually looks like: not prompt tuning, but playbook architecture.

Are you building agents with static prompts, or are you architecting self-improving context systems?

Reference: the ACE paper on arXiv (opens in a new tab)

Further reading

  1. Most agent failures are context problems, not prompt problems Prompt engineering is how you ask. Context engineering is what the model sees. Eight practices, the symptoms of a badly built context and a four-question audit for the window. 2 min readShort post
  2. Onboard your AI assistant like a junior developer Teams spend weeks onboarding a junior developer and minutes onboarding their AI assistant. Both need the same thing: context, standards, and patience. 2 min readShort post
  3. Reduce and delegate: two ways to keep an agent's context focused A context window is not infinite RAM. First cut what never belongs in it, then spread the rest across sub-agents and context bundles so long tasks keep their focus. 3 min readShort post

All 16 posts on Context engineering

ClientsCase study: A large DIY retailer in Germany

Test the constraint in your system

The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.

Training on this topic: Agentic Engineering for Developers and QA
See the assessment

Want to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check