What is SEIF
The mental model behind SEIF — workspaces, modules, evidence, and operating conduct.
SEIF turns ephemeral AI conversations into durable, verifiable work. Four ideas carry the whole model.
Workspace
A project directory with a .seif/ store: memory, keys, and governance records.
Module
A .seif file holding structured, replayable knowledge.
Evidence
A signed, timestamped record of an artifact, addressable by a public URL.
Conduct
The operating maxims every agent follows: settle, split attention, defer to a human.
Workspace
A workspace is any directory where you ran seif init. It owns a .seif/ store
that holds memory modules, signing material, and the governance ledger. In these
docs we refer to its root as $SEIF_CONTAINER_ROOT — it is wherever you chose,
not a fixed location.
Module
Knowledge lives in .seif modules: structured, human-readable files that AI
tools load at session start. A module records decisions, conventions, and
observations so that context is never lost when a chat window closes.
$SEIF_CONTAINER_ROOT/.seif/modules/decisions.seifEvidence
When you sign content, SEIF produces an Evidence URL. It resolves to a record containing the signature, the content hash, the timestamp anchor, and the classification. Evidence is verifiable by anyone, independent of the tool that created it.
Conduct
Every agent that speaks SEIF follows a small, signed set of operating maxims: how to settle on one considered answer, how to split attention between task and context, and when to defer to a human. The conduct has a control-theory lineage — a model agents once derived at session start — but it stands on engineering merit, not on the math; a pre-registered test refuted the idea that the mathematics itself improves quality.
For integrators
You are wiring SEIF into another system — an editor, an agent framework, or a provenance pipeline. Connect over MCP and the CLI, then verify.
Just-in-time expertise
SEIF makes the AI a just-in-time expert in its own capabilities, so the human never has to be. Surfacing a compact capability map lifted a cold model's verb-reach from 0.50 to 1.00 — measured.