HOW IT WORKS

Capture. Distill. Recall. Compound.

01

Capture

Every turn from any of six harnesses is captured over loopback — nothing leaves your machine.

02

Distill

Raw events become source-backed facts, entities, and skills, each carrying its provenance.

03

Recall

The right context surfaces before the next turn — hybrid lexical (BM25) + 768-dim semantic.

04

Compound

The Dreaming loop merges duplicates, prunes junk, and supersedes stale claims so memory sharpens.

27.5% avg.
reduction in token spend
1.4x avg.
faster llm responses
6+
harnesses supported

based on Hivemind benchmarks

SOURCE-BACKED MEMORY

Receipts, or it doesn’t count.

  • Every recalled memory shows its source — a file path or session id.
  • Every memory carries a score, so you can see how strong the match is.
  • Verified, source-backed claims are marked in green.
memory · verified ✓ verified
refresh handled server-side
auth/refresh-token.ts:42 personal 0.94
memory · recalled
staging uses the eu-west cluster
session 8f2a · 2026-06-19 personal 0.86
memory · consolidating dreaming…
consolidated overnight by the Dreaming loop
merged · 3 sessions personal 0.81
THE POLLINATING LOOP

While you sleep, it gets smarter.

The Dreaming loop goes back through everything learned, ties the pieces together, and wakes up sharper instead of noisier.

merged duplicate memories are tied together
superseded stale claims are replaced by newer ones
skillified recurring traces become reusable team skills

Private by default. Your store, your rules.

Your store, powered by Deep Lake

Memory lives in your own Deep Lake store — vector + columnar, inspectable.

Loopback only

The daemon binds 127.0.0.1:3850 and never opens a public port.

Versioned & inspectable

Every change is versioned; you can read, scope, and repair what it remembers.