The paper that finally unifies agent memory, skills, and rules
New arxiv paper called "Experience Compression Spectrum" takes 1,100 references across 22 papers and makes a claim worth pausing on: agent memory and agent skill discovery are the same problem, just at different compression levels.
Episodic memory compresses experience 5-20×. Declarative rules compress 1,000× or more. Between them sits skills — reusable procedures that are more abstract than a stored interaction but less abstract than a rule. The authors argue every existing agent memory paper and every existing agent skill paper is solving a point on this spectrum, usually without realizing the other community exists.
The more interesting claim is the "missing diagonal" — no current system moves adaptively across compression levels. Your agent either stores raw trajectories or learns symbolic rules. Nothing dynamically decides when an episode should become a skill or when a skill should become a rule. That gap is the real research frontier.
No code, pure survey. But if you're building an agent with memory, read this before picking a framework. It will save you from reinventing decisions the paper already catalogs.
Paper at https://arxiv.org/abs/2604.15877
← Back to all articles
Episodic memory compresses experience 5-20×. Declarative rules compress 1,000× or more. Between them sits skills — reusable procedures that are more abstract than a stored interaction but less abstract than a rule. The authors argue every existing agent memory paper and every existing agent skill paper is solving a point on this spectrum, usually without realizing the other community exists.
The more interesting claim is the "missing diagonal" — no current system moves adaptively across compression levels. Your agent either stores raw trajectories or learns symbolic rules. Nothing dynamically decides when an episode should become a skill or when a skill should become a rule. That gap is the real research frontier.
No code, pure survey. But if you're building an agent with memory, read this before picking a framework. It will save you from reinventing decisions the paper already catalogs.
Paper at https://arxiv.org/abs/2604.15877
Comments