Recent research proposes the "Experience Compression Spectrum" framework that unifies Memory, Skills, and Rules in LLM Agents on a single axis, arranged by data compression levels from experience. This addresses the problem where AI memory and skill discovery communities work separately despite solving common problems.
The Core Problem: Memory and Skills Working in Silos
Research from arXiv (2604.15877) found that AI memory and skill discovery communities work almost entirely separately, despite solving common problems (cross-community citation < 1%). Both communities lack a holistic view of agent learning systems. This problem stems from looking at components separately: Memory systems - Focus on storing and retrieving information Skill frameworks - Focus on discovering and using skills Rule systems - Focus on defining policies and constraints
Experience Compression Spectrum: Three Levels of Experience Compression
This new framework arranges agent learning on a single axis according to data compression levels from experience:
1. Episodic Memory (5-20x compression)
Stores raw experiences from interactions, such as complete conversations, actions, and outcomes. This is the lowest compression level that preserves the most information. Example: Agent remembers the entire conversation with a user about debugging code.
2. Procedural Skills (50-500x compression)
Extracts reusable skills from the agent's work history, such as general debugging methods, architecture design, or test case writing. Example: Agent learns a "skill" for checking dependency conflicts from solving similar problems multiple times.
3. Declarative Rules (1000x+ compression)
Summarizes into clear and concise rules, such as "Always use TDD" or "Never use global variables." This is the highest compression level. Example: Agent concludes that "Using interfaces reduces coupling between modules better than inheritance."
Analysis of 22 Major Systems and Discovery of "Missing Diagonal"
The research analyzed 22 major systems and found that: All systems operate at fixed compression levels; none support adaptive cross-level compression A phenomenon called "missing diagonal" emerged - no systems effectively link between different compression levels
Impact on Current AI Agent Systems
Problems Found in Current Systems:
- Memory overload - Agents store too much data without compression
- Skill redundancy - Multiple overlapping skills
- Rule inconsistency - Conflicting rules
- Learning inefficiency - Agents learn slowly due to lack of systematic compression
New Opportunities:
- Adaptive learning - Agents can adjust compression levels as needed
- Cross-level transfer - Transfer knowledge between different levels
- Efficient storage - Reduce memory usage with intelligent compression
Application to AI Systems Like Hermes
The concept of procedural memory from agent history can directly improve Hermes' knowledge management system: From conversations to skills: Transform conversation experiences into reusable skills Automatic compression: System selects whether to store as episodic, procedural, or declarative Cross-level linking: Ability to trace back from rules to skills and episodic memories
The Future of Experience Compression
This framework opens the door to: Human-like AI learning systems - Compressing experiences like humans Agents with "common sense" - From well-compressed rules Community collaboration - Memory and skill research working together Experience Compression Spectrum is not just theory but a practical framework that will help AI agents learn and develop more efficiently and sustainably. Original Source: Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents