The Knowledge Graph
The Kernel learns how memory connects, not just what it contains. Context becomes “the things related to what you asked” instead of “whatever matched a keyword” — and every connection is honest about whether you stated it or the graph inferred it.
Click a node to seed the Context Engine · drag any node to rearrange
Select a node in the graph to expand connected memory within the granted scope.
The design
Nodes are identity, capsules, people, and concepts. Edges — authored, knows, mentions, related-to — each carry a provenance tag. Nothing is connected anonymously.
self-asserted — you stated it (authored a capsule, named a relationship). derived — the graph inferred it (a capsule mentions a concept; two capsules share one). The two are never merged (Law 12).
The Context Engine expands from a seed through the graph, but only includes capsules whose dimension is in the permission scope. The graph may know a connection; the agent only receives what you permitted (Law 4). Connected and scoped.
Honest sharp edges
- Derived edges are heuristic. V1 infers concept mentions by keyword match over a small list — shallow, and it will both miss and over-connect. It is tagged
derivedprecisely so it is never trusted as fact, and can always be seen and pruned. Better extraction (embeddings, NER) is a later upgrade; the provenance discipline is what must not change. - Scale. This is an in-memory graph rebuilt from vault state — right for a personal vault, not millions of nodes. At scale, persist the graph and update it incrementally as events arrive (it pairs naturally with the CRDT event log: each
capsule.set/deleteis a graph mutation). - No auto-promotion. Nothing here ever turns a
derivededge into aself-assertedone. If you confirm an inferred link, that becomes a new self-asserted edge — a deliberate act, never automatic.