DEVOURED
Experience Graphs: The Data Foundation for Self-Improving Agents
Meta researchers propose treating agent search trees as 'experience graphs' stored in a database to enable cumulative agent learning.
What: Trellis is a proposed data foundation that stores agent search traces, rewards, and causal history as queryable state, allowing agents to resume sessions, reuse subtrees, and extract training data through SQL/graph/vector queries.
Why it matters: This signals a transition from viewing agent inference as disposable log-based processes to treating it as persistent, institutional data assets that support recursive self-improvement.
Deep dive
- Trellis manages agent state as a persistent, queryable Experience Graph (Causality, Executability, Reward).
- Architecture separates compute (stateless agents) from storage (durable database).
- Common operations like resume, reuse, and repair are treated as database access patterns rather than application logic.
- Enables 'vector-seeded graph expansion' to find and reuse high-performing past search trajectories.
- Training data extraction is performed via materialized views, eliminating the need for post-hoc scraping.
- Supports concurrency by managing tree statistics and backpropagation within the database transactional scope.
- Evaluation on 'KernelEvolve' showed 10x faster convergence and 52% lower token costs for optimized hardware kernels.
Decoder
- RSI (Recursive Self-Improvement): A process where an agent iterates on its own performance, using past outcomes to guide and improve future search and training.
- MCTS (Monte Carlo Tree Search): A search algorithm used in decision-making and AI that builds a search tree incrementally by balancing exploration and exploitation.
- Vellox: Meta's open-source, high-performance vectorized execution engine.
- DPO (Direct Preference Optimization): A technique to align LLMs to human preferences without requiring a separate reward model.
Original article
Full article content is not available for inline reading.
Explore all 65 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi