Weft

How Weft works

Memory in Weft has a shape: turns become claims, claims carry their receipts, and recall puts the right claim in front of the agent that needs it. Nothing is remembered that you didn't say.

Everything starts as a turn

A turn is one exchange with your agent — raw, timestamped evidence. Turns are honest about what they are: they expire. They are the substrate memory forms from, not the memory itself.

Durable material becomes a claim

What you designate as worth keeping crystallizes into a typed memory — a preference, a decision, a milestone, a fact — with provenance back to the turn it came from. Weft adds the provenance and context. You make the call on what's durable.

Reinforcement, not duplication

Mention the same thing again and the claim gains support instead of spawning a near-duplicate. Support counts and provenance chains are how a memory earns confidence — quietly, traceably, on the record.

Recall is semantic

Weft stores embeddings alongside text, so recall is a meaning-based search over what you saved — not a folder you have to remember. weft_prime assembles the context at session start; weft_recall pulls specific memories when asked.

Handoff closes the loop

End a session with weft_handoff and the next one — same agent, different agent, different machine — picks up where you left off. One command, no re-explaining.

This is user-directed memory with provenance, not autonomous curation. Weft doesn't silently summarize your conversations to pad a recall score. When it gains an autonomous classifier, it will be because it makes the core product better.

Watch a memory form

The audiobook example, five beats. Step through to see a turn become a claim — and come back when it's needed.

  1. Turn

    A turn arrives

    “I like listening to audiobooks on my commute, it’s about 45 minutes, so I prefer shorter chapters.”

    Every session is made of turns. Weft records the turn as raw, timestamped evidence — nothing has been decided about it yet.

  2. Parsed

    Facts surface — the turn stays a turn

    Structured facts surface from the sentence: the commute is about 45 minutes, and shorter chapters are preferred in the context of that commute.

    The turn remains a turn — raw, timestamped evidence with a lifetime. Nothing has been promoted, nothing dropped.

    • commute ≈ 45 min
    • prefers shorter chapters
    • context: commute

  3. Claim

    A claim forms — with receipts

    What’s designated as durable crystallizes into a memory: a typed claim with provenance back to the turn it came from. Not a summary Weft invented — a claim, with receipts.

    preference.audiobook-chapter-length source: turn #412 · 2026-09-29 08:14

  4. Reinforced

    Reinforced, not duplicated

    Weeks later you mention audiobooks again. Weft strengthens the same claim instead of spawning a near-copy — the support count goes up, and the provenance chain grows with it.

    preference.audiobook-chapter-length source: reinforced · turn #587

    support_count +1 — the claim deepens instead of duplicating.

  5. Recalled

    Recalled when it matters

    “Find me an audiobook for my commute.”

    Weeks after it was saved, the claim appears in context — because it was saved, and because it can be found.

    Nothing was remembered that you didn’t say. Nothing useful was lost.

A scripted visualization — no live calls to any Weft instance. It shows what the lifecycle looks like, not a running system.

Try it on your own machine

The walkthrough above is five beats of a real lifecycle. Yours starts with one install command.

Get started