Ontology Graph
The Ontology Graph is the connective layer of your Personal Brain. It turns your tree-shaped LLM Wiki and your event timeline into a richly linked graph, so agents can follow relationships, not just look things up.
Your Memory Events and your LLM Wiki already hold a lot about you — but each is limited by its shape. The wiki is a tree; the events are a timeline. Neither can express the web of relationships that actually connects your world. The Ontology Graph adds that web: a typed, directed graph laid over your wiki nodes and events, grown by an agent that explores your memory the way you would.
Why a tree isn't enough#
The LLM Wiki is a hierarchy. Every node lives at a path — wiki://you/entities/acme, wiki://you/preferences/database — and every node has exactly one parent. That makes it easy to browse and cheap to store, but a tree can only encode one kind of relationship: containment. One path in, no cross-links, no cycles.
Your actual knowledge is not shaped like that. The same entity turns up in a dozen places; a preference is tied to a project, a project to the people on it, a decision to the event that triggered it. Those are sideways links a tree structurally cannot hold. The event timeline has the opposite gap: it is perfectly ordered but flat — it knows when things happened, not how they relate. The Ontology Graph exists to supply exactly the edges both are missing.
A tree is just a graph with one rule: a single parent per node. Drop that rule and you get a graph — any node can link to any other, with typed, directional edges and cycles. That freedom is precisely what's needed to model a person's world, where everything relates to everything.
Nodes and edges#
The graph is built from what you already have. Its nodes are your LLM Wiki entries (profile, preferences, entities, milestones, cases, patterns) and your Memory Events. Its edges are the typed relationships between them — discovered, not declared.
Every edge has a type and a direction — authored, attended, caused, prefers — so the graph encodes what a relationship is, not merely that one exists.
The same person, project, or tool mentioned across many nodes and events is merged into one node, so all of its relationships converge instead of scattering.
Each edge records the wiki node or event it was derived from, so any relationship can be traced — and trusted — back to your own memory.
Built by agentic exploration#
The graph is not produced by a single extraction pass over your text. It is grown by an agent that explores your memory — walking the wiki tree and the timeline, reading nodes at increasing detail, and proposing relationships it can then check. Because it is explorable, the graph is also incremental: it deepens where your memory is dense and stays cheap where it is sparse, expanding on demand as new memory arrives.
The agent starts from a node — an entity you mentioned, or a fresh event — and walks outward along the wiki hierarchy and the timeline, gathering its neighborhood.
Reading the connected nodes and events, it proposes candidate edges — entity↔event, entity↔entity, event↔event (temporal or causal), preference↔project — each with a type and a direction.
Every candidate is checked against the underlying memory. Links the evidence doesn't support are dropped, so the graph stays grounded rather than speculative.
Surviving edges are written into the graph, each carrying the wiki node or event it came from — so every relationship is traceable back to its source.
Relationship-aware retrieval#
The payoff is multi-hop reasoning. A question like "what did I decide about pricing the last time Acme came up, and who was involved?" can't be answered from a single node — it has to hop from the entity to its events, and from those events to the decisions and people attached. That traversal is exactly what the graph makes possible, and it is one of the three concurrent searches an agent runs during retrieval. The wiki tells you where things live; the graph tells you how they connect — and together they let your three memory systems reason as one.
The curated, trustworthy core of your memory, and why it holds up.