Compiled Memory: Not More Information, but More Precise Instructions for Language Agents

TL;DR

This paper introduces Atlas, a memory system that compiles task experiences into instruction rewrites for language agents, improving performance without traditional storage or fine-tuning.

Contribution

Atlas offers a novel approach by distilling experience into instruction prompts, enhancing agent behavior without fine-tuning or retrieval-augmented generation.

Findings

Improves GPT-4o token-level F1 by +8.7pp on CUAD.

Enhances HotpotQA joint F1 by +3.16pp.

Task-specific compiled knowledge benefits different models.

Abstract

Existing memory systems for language agents address memory management: how to retrieve and page more information within a context budget. We address a complementary problem -- memory utility: what experience is worth keeping, and how it should change agent behavior. We present Atlas, a memory kernel that compiles accumulated task experience into an agent's instruction structure -- without fine-tuning, RAG, or human intervention. Memory is distillation, not storage; delivery is instruction rewriting, not context injection. Facts extracted from agent failures and successes are verified through a three-step promotion gate and delivered by rewriting the agent's system prompt with learned sub-bullets. On CUAD contract analysis, the evolved prompt improves GPT-4o token-level F1 by $+8.7$pp and precision by $+12.5$pp. On HotpotQA multi-hop QA, joint F1 improves $+3.16$pp. An ablation isolates the mechanism's defining property -- the training signal constraint: the evolved prompt learns exactly what it is taught, and nothing more. Applied to Claude Sonnet~4.5 using the same evolved prompt -- compiled from GPT-4o errors, unchanged -- joint F1 improves $+2.31$pp, with gains concentrating where Claude's stronger baseline leaves the most room -- confirming that the compiled knowledge is task-shaped, not model-shaped.