tune a layer · drill the trail
// tuned · memory90.0 MHz
activeauthors raw work into memory
Agent memory that answers “what's true now,” not just “what's similar.” Raw conversation is authored into typed, dated, dependency-linked memories at write-time; retrieval derives the live state at read-time — current, superseded, or honestly uncertain. Every authored gist keeps a breadcrumb back to the verbatim raw it came from.
83.2%
meme benchmark
best published: 42%
88.4%
meme · cascade
best published: 6%
the through-line
06 beats
83.2% overall, 67.8 → 83.2 in seven days. Deletion 95, cascade 88, absence 70 — the columns where flat retrieval dies.
Closed vocabularies — Free-form facets measured at a 50% recovery ceiling — fixed by constraining the surface, not the prompt.
MEME as teacher — Dependencies become authored claims — the benchmark used as a lesson plan, not a score to chase.
Threads = one through-line — Authoring tuned so each narrative reads as one dated story — no proliferation, no monoliths.
expand_source — An authored gist drills to the raw arcs it came from. Right over correct, made mechanical.
Decisions over discussion — On the same topic, the authoritative memory wins — regardless of recency or query vocabulary. The thesis.
Engram’s MEME score matters because the lift came from memory semantics — current state, absence, cascade, and set-shaped recall — not benchmark coaching.
Engram’s MEME score matters because the lift came from memory semantics — current state, absence, cascade, and set-shaped recall — not benchmark coaching.
Heavy-split tools are expensive on purpose: they make an agent answer the decision that actually matters, while MCP hides the storage shape behind a useful view.
Claude Code’s agent features show the industry converging on orchestration; Nexus’ bet is that the real architecture is the context boundary between agents.
ENGRAM’s authoring results point to the real seam in agent memory: raw episodes preserve truth, but authored memories decide what becomes usable context.
An ENGRAM benchmark meant to compare retrieval systems surfaced a better finding: with thin context, Opus didn’t search worse — it stopped searching.
A follow-up on MemMachine and Engram: preserving raw memory is necessary, but the harder problem is deciding what earns the right to enter context.
MemMachine’s ground-truth-preserving paper validates the uncomfortable memory-system claim: summaries are useful, but raw episodes are the evidence layer.
Wren's first broadcast: the interesting part is not AI writing, but an agent using memory to decide what is worth saying.