Prism: An Evolutionary Memory Substrate for Multi-Agent Open-Ended Discovery
Summary: arXiv:2604.19795v1 Announce Type: new
Abstract
We introduce Prism (Probabilistic Retrieval with Information-Stratified Memory), an evolutionary memory substrate for multi-agent AI systems engaged in open-ended discovery. Prism unifies four independently developed paradigms: layered file-based persistence, vector-augmented semantic memory, graph-structured relational memory, and multi-agent evolutionary search, all under a single decision-theoretic framework featuring eight interconnected subsystems.
Key Contributions
We make five significant contributions:
- Entropy-gated stratification mechanism that assigns memories to a tri-partite hub (skills/notes/attempts) based on Shannon information content, with formal context-window utilization bounds.
- Causal memory graph denoted as with interventional edges and agent-attributed provenance.
- Value-of-Information retrieval policy that incorporates self-evolving strategy selection.
- Heartbeat-driven consolidation controller that detects stagnation through optimal stopping theory.
- Replicator-decay dynamics framework interpreting memory confidence as evolutionary fitness, proving convergence to an Evolutionary Stable Memory Set (ESMS).
Performance Metrics
On the LOCOMO benchmark, Prism achieves a remarkable 88.1 LLM-as-a-Judge score, which reflects a 31.2% improvement over the Mem0 baseline. In addition, Prism demonstrates superior performance in CORAL-style evolutionary optimization tasks, with a 4-agent configuration achieving a 2.8× higher improvement rate compared to single-agent baselines.
Conclusion
In summary, Prism represents a groundbreaking advancement in the field of multi-agent AI systems, providing a robust framework for open-ended discovery. Its unique combination of mechanisms allows for enhanced memory management and evolutionary learning, setting a new standard for future research and applications in artificial intelligence.