Notes from building Metatron
Why coding agents need a memory of your team's decisions — and what we're learning as we build the layer that gives them one.
Context Inheritance: the presentation
The slides for our AgenticDev @ ASE 2026 talk in Munich. Browse the presentation, open it fullscreen, or download the PDF. Includes all 18 talk slides and 6 backup slides.
DEMOThe tests passed. The payment ran twice.
Both coding sessions wrote passing tests. Only the one with stored team decisions handled a payment retry correctly. Watch the code, the comparison, and the setup in two short demos.
RESEARCHRipwire and Metatron: code intelligence meets repository memory
Ripwire derives a map of the repository today. Metatron preserves what engineers and agents learned yesterday. Why coding systems may need both derived and inherited context—and why Ripwire notes blur the boundary.
RESEARCHGraft, Metatron, and the two kinds of context coding agents need
Graft helps agents reconstruct what a codebase contains. Metatron helps them inherit what previous engineers learned. A technical comparison of structural code intelligence, durable engineering memory, and why the ideal repository context stack may need both.
ENGINEERINGVerification contracts: teaching your repo how to prove itself
Repositories record how to think about a codebase but never how to prove a change works — so agents reinvent the run-and-verify steps every session. A verification contract is a git-tracked artifact that captures them, plus a curated Failure Means section that routes a red check to the subsystem at fault. Authored by the agent that just built the feature; run only by a human or CI. Shipped in 0.13.0.
RESEARCHContext inheritance: who wrote the context decides whether it helps
Give a coding agent your repository's context and sometimes it transforms the agent, sometimes it does nothing. A pre-registered SWE-bench study found why: benefit depends on who wrote the context versus who reads it. A weak model doubles its localization from frontier-written context (+26pp) and gains nothing from its own. Context flows downhill.
ENGINEERINGSmall models don't follow policies. They follow recipes.
Metatron only works if agents actually read your decision files — so we measured it. A prose consult-first contract: 2/20 compliance. The same contract as a numbered recipe: 20/20. Plus the invisible-AGENTS.md bug and the one-line CLAUDE.md bridge. Shipped in 0.12.0.
WORKFLOWFiles-first: run Metatron with no server at all
Git as the source of truth: one command onboards a repo, any LLM authors candidate decisions as plain files, and promotion is a reviewed git mv. We dogfooded it on Metatron itself — and it caught a bug in our own importer on day one.
INTEROPMetatron speaks the Open Knowledge Format
Your codebase's decisions just left the database. Metatron can now mirror them into plain markdown in your own git repo — a valid Open Knowledge Format (OKF) bundle. Auditable in git, editable with your agent, portable to anything that reads OKF.
ORIGINWhy I built Metatron
A reasonable refactor silently broke my payment funnel for a day. The fix was one line — but the real cause was a decision that lived nowhere an agent could read it. That's the layer I set out to build.