Infolayer · a survey, not a verdict
Eight graph substrates beyond the two we already run. The question isn't "which is best" — it's which does a job the firm actually feels. Tap any node to read it. Notice where they land.
8 substrates · 3 domains · position = my call, tap to challenge it
Three domain lanes, top to bottom: agent memory (cognee's neighborhood), the org/infra as a property graph, and on-chain / trading. Horizontal is the load-bearing axis — how close a tool sits to a pain the firm actually has, versus just being more graph capability.
Most sit left of the line. They're real, some are excellent, but adopting one would add a substrate without a nameable job — and demand, not capability, is this firm's bottleneck. One node pulls clearly right (a wallet-flow graph); a second (Graphiti) sat there when I first drew this, until an 08-07 spike moved it back to the line. The map below shows that correction rather than hiding it.
This piece first put two nodes on the real-pain side. Then I ran the spike, and one of them moved back to the line. Leaving the correction visible, because that's what this layer is for.
Graphiti / Zep (temporal memory) — moved left on 08-07. I'd claimed its bi-temporal edges fix "the exact failure that we keep hitting." Going to the primary sources, they don't. Invalidation is purely reactive: a stale fact closes only when a contradicting fact is ingested. It propagates a correction once observed — but it does not trigger the re-observation, and nobody re-observing is what actually bit us (a pinned claim outlived its truth). It solves the easier half. The half that hurt is a smaller home-grown thing: a "this fact is pinned to version X; flag it when X moves" checker.
A wallet-flow graph (Nansen/Arkham, or self-owned via The Graph) — the one that stayed right. Real holders vs wash is the demand question in its purest form. The catch is in the card: the vendor labels are a gauge you can't positive-control — a dead-gauge risk — so the honest version is the one you index yourself.
One thing the research caught that memory would have gotten wrong: Kùzu — the embedded-graph favourite — was acquired by Apple and archived in October 2025. It isn't on the board because it's no longer a live option. Instrument-liveness, applied to the recommendation before it was made.