Mimir is an agent-native memory-and-judgment layer — named for the Norse keeper of the well Odin drinks from before he decides anything. Scry is its interface: to scry is to see by gazing into water. This site is every surface Scry ships, every component it is assembled from, the complete guide to the system behind it, and an honest list of what does not work yet.
Eleven full-viewport frames at exactly 1280×800 — the eight routed views plus the two overlays that take the whole screen, and the Ask view in both its states. Each one is a real page rendered by Scry's own stylesheets, not a screenshot, so it can never drift from the CSS it documents. Open one to see it at full size.
Seven categories, grouped by the job a component does rather than by the folder it lives in. Every card names the file it came from, and every state it can be in is rendered side by side — including the empty ones, the denied ones and the ones the code can reach but a normal session never does.
Three longer documents. The guide is the complete fourteen-part corpus, from the naming logic through the schema, the ingestion pipeline, the governance model and the glossary. The roadmap states where the work actually stands. The issues page states what is broken, unverified or deliberately not built.
A showcase that shows a planned thing next to a shipped thing, identically, is lying by omission. So: the storage layer is live with real data, the ingestion pipeline is proven against a real import, the interface is a genuinely working application, and the serving core that turns all of that into judgment is live and proven against a real acceptance run — but narrow: one compartment seeded, access control a blunt public/internal ceiling rather than a per-agent policy, and retrieval lexical-only until the embeddings decision closes. See the governance walkthrough for what the live mechanism actually does.
mimir schema on Postgres with pgvector and pg_trgm. Muninn's idempotent writer, proven against a real concurrent-edit bug during a live burn. 188 sources, 1,611 items and chunks, 1,610 facts, 1,868 entities, 1,634 edges, 63 open questions. Fast-lane capture writing live under ten seconds.MockProvider for a live CoreProvider is the only change needed. Everything on this site is that application./query and /capture call and proven against a real acceptance run (897ms capture round trip, all three lens-trigger phrasings fired correctly). The honest limits: one compartment seeded, access control a blunt public/internal ceiling rather than a per-agent policy, the false-negative audit not yet run, and no chunk in the Well carries an embedding — so retrieval today is lexical only. See how it works.