Skip to content

Home

Jump to bottom
Yuriy Orlov edited this page Aug 27, 2026 · 4 revisions

KAUT — Knowledge Actualization Under Trust

Self-maintaining, AI-first documentation for your project. KAUT is a knowledge layer that makes undocumented, weak-context code legible to AI agents: what the system does, how it is structured, and why — with every fact bound to its sources, anchored to a commit, and freshness-checked by pure git computation on every read.

Standalone by design: one git clone, zero dependencies, no services — connect the bundled MCP server to any MCP-capable harness (or call the CLI from any skill) and everything works. The sibling TAUT orchestration framework integrates KAUT natively (TAUT drives agents, KAUT is what they know), but that pairing is optional — KAUT needs nothing besides Node and git. The store format is an implementation of the vendor-neutral Open Knowledge Format (OKF) v0.2 — portable typed markdown, no lock-in.

Start here

Page What it answers
Why KAUT Why this is not another memory system, not RAG, not an LLM wiki — and what it actually is
Getting Started Install, first store, first lookup — under five minutes
Core Concepts Verdicts, trust tiers, altitude, layers, the write gate — the vocabulary everything else uses
Connecting Your Project How knowledge binds to your code: sources, anchors, layers, maps, multi-repo workspaces
Using With Agents The step that makes it real: how your agent learns to consult the store and write back
MCP Server Plugging any MCP-capable harness or orchestrator into KAUT
Maintenance Loop How knowledge stays fresh: touched / refresh / drafts / review / telemetry
FAQ Short answers to the questions everyone asks
Troubleshooting Verdicts, doctor output, exit codes — what they mean and what to do

The two-minute version

  1. Install: git clone the engine next to the repositories it will serve (zero dependencies, Node ≥ 20).
  2. Run node kaut/kaut.mjs setup: pick the data folder (live data — the engine only ever adds to it), select the sibling repos from the scanned list, let bootstrap run. Stores are created outside your repositories — they are never touched.
  3. Wire your agents (Using With Agents): register the MCP server with your harness and paste the ~15-line knowledge contract into CLAUDE.md / AGENTS.md — that is how an agent learns the store exists and when to consult it.
  4. Work as usual: agents look up before re-exploring code; every answer arrives with a freshness verdict that routes trust — healthy means usable, stale means "re-check, here is exactly what changed".
  5. Knowledge accumulates: verified operational facts land at agent tier through the write gate; judgment-tier knowledge (decisions, domain semantics, contracts) queues as drafts for your one-sitting review.

The wiki gives you the guided path; the repository docs are the reference — docs/HANDBOOK.md (full mechanics), docs/OPERATIONS.md (operator depth), docs/MCP.md (MCP reference), SCHEMA.md (the normative data contract).

Apache-2.0 · Copyright 2026 Yuriy Orlov · yuriy.orlov@undertrust.dev

Clone this wiki locally