Memory, Skills, and Rules Aren't Three Different Systems. They're the Same System at Different Compression Levels.

Memory, skills, and rules aren't three different systems. They're the same system at different compression levels.

A new paper surveyed every major agent knowledge system — Mem0, Voyager, SkillWeaver, all of them — and formalized something builders have felt but couldn't articulate.

Level 1 — episodic memory. "User requested Q3 revenue analysis via SQL. Preferred tabular format." 15× compression. Useful for that user. Doesn't transfer.

Level 2 — procedural skill. "DATA_ANALYSIS: confirm source → select tool → present in preferred format → verify." 500× compression. Transfers across tasks.

Level 3 — declarative rule. "Always verify computed results against source data." 1,000×+ compression. Transfers across everything.

Agents with Level 2 skills outperformed agents with Level 1 memory by 68.5 percentage points on the same benchmark. Not marginal. A different category.

The scalability math: 1,000 episodes at Level 1 = 500K tokens searched every decision. Compress to Level 2: 5K tokens. Level 3: 500 tokens. A 1,000× reduction compounding across thousands of daily decisions.

The problem: every existing system operates at exactly one level. Memory systems never compress upward into skills. Skill systems never consolidate into rules. The authors call this the "missing diagonal."

Auto-generated skills provided zero benefit. Curated skills improved performance by 16 points. Compression without quality control is compact noise, not compact knowledge.

If your agents store everything but compress nothing, retrieval costs grow linearly forever.