Weft saves what you tell it to save.

What is Weft?

Weft is a persistent brain you share with your agents. The default alternative is a set of flat files your agent reads and writes. That works for a while, until it doesn't:

• Memories stay locked to one repo or one platform — another agent can't read them.
• Retrieval is limited to grep, so finding one memory can mean burning your agent's context reading everything.
• Nothing tells the agent apart what matters in every session from a rule you wrote for a bug you fixed nine months ago.
• Every new session starts from zero, handing the agent the same list of files again.

Weft treats memory as a first-class data system. Multiple agent platforms read and write the same brain. You can hand a session off to a fresh agent with one command. Same agent, different agent, different machine — it doesn't matter. Weft is built for your projects, not for a single platform.

How memory is formed

A turn arrives. What you designate as durable becomes a claim with provenance back to the turn it came from. Later mentions reinforce the claim instead of duplicating it. Recall puts it in front of the agent that needs it.

See the five-beat walkthrough →

The numbers, plainly

Weft doesn't benchmark well — because it isn't designed to remember things the way a benchmark tests. For transparency, we ran it through LongMemEval-S (multi-session chat memory, ~40 sessions per question) using the turn tier: the memory representation that works best here.

On the full 500-question split, judged accuracy was 67.54% — one paid run, scored with LongMemEval's official answer-check prompt, nothing hidden. The weakest question types (multi-session, knowledge-update, single-session preference) are named as the first post-launch improvement targets. The numbers measure the system as it actually ships, not a benchmark-tuned configuration.

Full results, per-type breakdown, cost ledger, and the harness to reproduce it — all in the repo: docs/benchmarks.md.