|
Build, evaluate, and integrate long-term memory for self-evolving agents. |
A scalable, end-to-end trainable latent-memory framework for 100M-token contexts. |
Methods are production-ready memory architectures that give agents persistent, structured long-term memory. Each can be used standalone or composed together depending on your use case.
|
A self-organizing memory operating system inspired by biological imprinting. Extracts, structures, and retrieves long-term knowledge from conversations — enabling agents to remember, understand, and continuously evolve. |
A hypergraph-based hierarchical memory architecture that captures high-order associations through hyperedges. Organizes memory into topic, event, and fact layers for coarse-to-fine long-term conversation retrieval. LoCoMo 92.73%. |
Benchmarks are designed as open public standards. Any memory architecture or agent framework can be evaluated under the same ruler.
|
Three-layer memory quality evaluation: factual recall, applied reasoning, and personalized generalization. Evaluates memory systems and LLMs under a unified standard. |
Agent self-evolution evaluation — not static snapshots, but longitudinal growth curves. Measures transfer efficiency, error avoidance, and skill-hit quality through controlled experiments with and without evolution. |
