Understanding Graph

Think git for AI cognition. Every mutation flows through `graph_batch` with a commit message, wrapped in SQLite transactions that roll back atomically on failure. Nodes are never deleted, only superseded via typed edges. The API exposes `graph_skeleton` for orientation, `graph_semantic_search` for retrieval, and `graph_history` to read what other agents did recently. Multiple agents coordinate through stigmergy: they read the graph, add reasoning nodes (tensions, surprises, decisions), and leave traces for the next session. Cross-project references via `graph_add_reference` let you build hierarchies across projects. Ships with Claude Code skills, optional embedding search via transformers, and a 3D web UI. Reach for this when you need agents to remember not just facts but how they arrived at conclusions.