Theo Valmis

I started Mneme HQ because of a problem I kept hitting when I used AI coding tools on production systems. The bigger the codebase got, the less I could trust what the tools wrote.

The code itself was fine. A model would write a working function in seconds and follow whatever I told it in that session. In the next session it would reach for a pattern the team had rejected months earlier, or ignore a constraint I had explained the day before. I kept re-explaining the same architecture decisions to a tool that treated them as optional.

Why I built this

My background is in digital analytics, AI systems and large-scale technical delivery. Over more than 20 years I have worked with startups and with enterprise teams at the organisations below, on platforms where a team that ships fast and breaks its own architecture pays for it later.

When AI coding tools became part of my own workflow, I watched them speed up implementation and erode the architecture at the same time. Nothing checked generated code against the decisions the team had already made until a reviewer caught the problem. I built Mneme HQ to run that check earlier.

What Mneme HQ does

Mneme HQ is the open-source decision and control layer for agentic software development. Architecture is the first decision domain. You record your project's architectural decisions, constraints, approved patterns and anti-patterns as structured governance records in the repository. Mneme retrieves the ones that apply to the task in front of the agent and evaluates the output against them before it reaches review.

You can use it to enforce:

• Architecture decisions
• Coding standards and tooling preferences
• Anti-patterns and rejected approaches
• Project-specific implementation rules

The same records work across Claude Code, Cursor, GitHub Copilot and Codex CLI, so a decision holds whichever agent writes the code.

Why this matters now

AI coding agents write code faster than anyone can review it. An agent has no durable way to know which human decisions govern the work it is doing, so approved architecture becomes context it may or may not follow. Patterns diverge between sessions, contributors and tools, and reviewers catch the drift after the fact, if they catch it.

Generation speed stopped being the bottleneck a while ago. The question I care about is whether the output follows the decisions the team has already made.

---

How to engage with the project

The fastest way to understand Mneme is to read the architectural governance article, watch the interactive demo, and skim the benchmark methodology.

If you're an engineering leader: the CTO page covers the strategic case, the platform engineering page covers the rollout pattern, and the standards landscape covers how Mneme aligns with NIST CAISI, MCP, and AGENTS.md. I track that work because long-term governance infrastructure cannot be vendor-coupled, and because the regulatory direction matters for any team operating in healthcare, finance, or public-sector environments.

If you're an individual engineer: install with pip install "mneme-hq>=0.10.0" && mneme init, wire the Claude Code hook from the integration guide, and run mneme check against your repo. Everything is open source under MIT, with no signup and no waitlist.

Talks and publications

I spoke at O'Reilly's Software Architecture Superstream: Building Architectures for Agents (and Humans), hosted by Neal Ford, on September 29, 2026. My session, Architectural Guardrails for AI Agents: Turning Decisions into Enforcement, covered how to move architectural decisions out of human-readable documentation and into a governance layer agents can use: the difference between retrieval and applicability, between guidance and deterministic enforcement, and how to keep architectural intent consistent across coding tools.

I wrote Architectural Guardrails for AI-Generated Code for O'Reilly Radar (August 2026).

I contributed to KodeKloud's The Definitive Guide to AI for DevOps (2026), written by Jennifer Riggins from interviews with 29 practitioners across 24 organizations, alongside engineers from AWS, Google, Oracle, Nvidia, Honeycomb and Netlify. My contribution appears in section 06, Guardrails, not gates, and argues that architectural intent has to become machine-readable, enforceable, and continuously available wherever development happens.

Connect

I build Mneme HQ in the open. If you are working on governance or decision enforcement for AI-assisted development, write to me.