Kumiho: Graph-Native Memory for AI Agents

🔑 Enhanced Key Takeaways

• •Kumiho offers a cloud service at kumiho.io alongside open-source Python SDK, MCP memory plugin, and benchmark suite on GitHub.
• •Dream State is an asynchronous consolidation process that enriches memories by reviewing for staleness, adding semantic tags, creating relationship edges, and deprecating superseded facts overnight.
• •Kumiho includes integrations like an OpenClaw plugin with zero-latency prefetch recall and two-track consolidation (threshold and idle), plus a Claude skill for persistent memory in AI sessions.

🛠️ Technical Deep Dive

• •Dual-store model uses Redis for short-term working memory buffer and Neo4j for long-term graph storage.
• •Hybrid retrieval combines fulltext, graph traversal, and vector search; prospective indexing generates LLM-based future-scenario implications at write time.
• •Client-side LLM reranking improves accuracy (e.g., GPT-4o-mini to GPT-4o boosts from ~88% to 93.3%); Principle 6 stores minimal metadata in cloud graph while keeping raw content local for privacy.
• •OpenClaw plugin features zero-latency prefetch (parallel recall during response generation), two-track consolidation (threshold: 20 messages; idle: 300s), and creative_recall for versioned agent outputs with krefs.
• •Local mode recommended with cloud HTTPS API key option; Dream State uses LLM auth for scheduled maintenance without separate keys.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline