Knowledge Graph · Field notes · Part 3 of 3

Knowledge Graph, Part 3: A living interface

Extraction, design tokens, better answers from the graph, and the Sci-Fi Labs frame for a sovereign memory surface — still inside a 16GB shared-memory budget, still offline.

Parts 1 and 2 made the canvas real and the queue survivable under a hard ceiling: the entire stack in about 16GB of shared memory, with offline as the proof the model is actually local. Part 3 is where the surface became intentional — component extraction, semantic tokens, and graph answers that keep same-day photos and notes together so the place you travel through earns trust.

Design tokens and chrome

oklch semantic tokens replaced ad-hoc hex. Message bubbles, thinking blocks, tool summaries, and stream cursors shared one language. ProcessingDock moved to the bottom-right, learned to poll the API, hide finished work, and offer reprocess on failure. Photo nodes stopped jumping time buckets: creation waited until EXIF arrived.

Why the graph answers better

Direction is wasted if asking “what did I do last week” returns a shallow summary. Date queries stopped relying on slow LLM keyword extraction; NLP shortcuts and post-filters kept same-day media together. Notes got their own plane kind. MCP tools could save and query; the model was instructed to actually call tools when it claimed to.

Sci-Fi Labs frame

Knowledge Graph sits with Where is Paul? and Musical Cubes under one thesis: spatial products for web, mobile, and XR. Memory as a place is the product constraint — not a feature list bolted onto a chat window, not a stack that only works when you rent more RAM in the cloud, and not an “AI” that dies the moment the network drops.