Research at HumemAI

Hybrid retrieval (relationship‑aware + vector): Combine semantic embeddings with relationship‑aware queries over graph‑structured representations for accuracy, robustness, and explainability—without forcing users to learn query languages. This enables robust agentic planning and tool use.

Ontology and graph construction: Use LLMs and ML pipelines to extract entities/relations, induce schema, and continuously evolve domain‑specific ontologies from raw data and interactions.

Memory management policies: Learn what to store, summarize, link, and forget using reinforcement learning and heuristics guided by task performance and retrieval quality—critical for long‑lived agents.

Query planning with LLMs: Plan and execute mixed SQL/NoSQL and relationship/graph operations, ground generation, and verify answers against the knowledge graph layer—foundations for agentic reasoning and action.

We publish, build open‑source tooling, and validate on real workloads so the results transfer directly into product capabilities. Collaborate with us.