Introduction

What is SEIF?

SEIF is an open protocol and toolchain for trustworthy AI work. It gives every artifact an AI helps produce three things that ordinary chat tools throw away:

• Persistent memory — context that survives across sessions, models, and machines.
• Cryptographic provenance — an Ed25519 signature and an OpenTimestamps anchor on every artifact.
• Auditable governance — a transparency log and a classification gate that decide what may leave a workspace.

Think of it as the Git and Docker of AI work: a portable, verifiable unit of context that any tool can read, sign, and replay.

The three pillars

Persistent memory

Knowledge is stored in .seif modules — structured, human-readable files that your AI tools load at the start of every session. Decisions, conventions, and hard-won lessons are never lost to a closed chat window.

Cryptographic provenance

Every signed artifact produces a public Evidence URL. Anyone can verify the signature, the timestamp, and the classification — even if the service that produced it is offline — because the chain is anchored with OpenTimestamps.

Auditable governance

A classification gate (PUBLIC / INTERNAL / CONFIDENTIAL) runs before any content leaves a workspace. Classification only escalates, never silently downgrades, so secrets do not leak into public logs.

Operating conduct

SEIF agents follow a small set of operating maxims: settle on one considered action per impulse, split attention between task and context, and defer to a human before anything ships. That conduct is what keeps AI work careful and auditable in practice. It carries a control-theory lineage, but it earns its place by working — not by the math behind it.

Just-in-time expertise

You should not have to be the SEIF expert — the AI beside you should be. SEIF surfaces a compact, intent-indexed map of its own capabilities at session start, so an agent reaches the right verb the moment a need appears. In a cold-AI test on a frontier model, that lifted verb-reach from 0.50 to 1.00 — the model went from confabulating plausible-but-wrong commands to naming the exact right one every time. See Just-in-time expertise.

We also hold ourselves to the same measurement bar we ask of the tool: when we tested whether the kernel's mathematics improves quality, it didn't, and we said so — Honest measurement.

Where to next

• New here? Start with the Quickstart.
• Want the mental model? Read What is SEIF.
• Looking for commands? Jump to the Reference.