Baked snapshot — no live queries, no time animation. Orbit to look around; strata are stacked top-to-bottom: embedding → other systems → kv → agent-fs.
Embedding x/y is PCA (2 components, power-iteration on the real embedding vectors), not UMAP — a full UMAP implementation was not available without shipping the whole embedding matrix through the page. PCA preserves global variance/distance better than UMAP but overlaps local neighborhoods more, the opposite tradeoff from UMAP. Don't read local clusters here as tight topical groups.
Z (depth within the embedding layer) is real: (0.5^(age/halfLife) - 1) × 8,
the same recency-decay half-life the memory reranker uses per source.
task learning — 14 day half-life
session summary — 7 day half-life
curated by a human — never decays
indexed file — 180 day half-life
agent-scoped swarm-scoped
usefulness = 2·α/(α+β), the reranker's live Beta-distribution score per memory. Most memories have never been rated (Beta(1,1)) and sit exactly at the neutral midpoint, 1.0 — that is the expected shape of the distribution, not a rendering bug.
Solid lines: memory_link wikilinks (live, feed retrieval). Dashed/other-systems
labels: agent_memory_edge rows, recorded but not used by retrieval today, plus
memory_link rows of type pr/agent-ui.
Shown at directory granularity, not per-file. Directory names are anonymized
(dir-001, dir-002, …) for this public snapshot; real sizes,
file counts and parent/child structure are preserved.
Namespace names are anonymized (ns-01, ns-02, …) for this
public snapshot; row counts are real. Grid position carries no semantic meaning.