Product Manager and systems architect building governed AI agents, local-first tooling and agent workflows.

github

Jérémy Grimonpont • Peerlist

<h3>Cortex gives memory to the AI agents you already use.</h3><p>You work with Claude Code, Codex, ChatGPT. Every session starts from nothing: you explain the context again, you repeat decisions already made, you redo trade-offs already settled. The model changes, and everything you built with the previous one stays behind.</p><p>Cortex captures the work you carry out in those tools, builds an experience from it, and returns the relevant context when they need it. It replaces nothing. You keep talking to your agents exactly as you do today, and Cortex runs underneath.</p><h3>What it keeps</h3><p>Remembering is not rereading. Cortex does not replay your old conversations. It extracts what carries forward: decisions and the reasons behind them, constraints, questions left open, what worked and what failed.</p><p>Every item keeps its provenance, its scope and its state. A conclusion established six months ago may have been contradicted since; it stays readable, but it loses the right to influence the present. Something learned on one project does not spread to the others on its own. Archiving is reversible, forgetting is not, and that last gesture asks for a written confirmation.</p><p>Work is organised into workspaces and topics. A topic carries its own memory. What follows you from one topic to the next, your methods, your constraints, the way you work, lives separately.</p><h3>You decide what goes in</h3><p>A session enters Cortex only if you allow it, and you can take that permission back. A learning Cortex has inferred waits for your approval before it can weigh on anything; you can then dispute it, remove its influence, or erase it.</p><p>Three levels of control: bring everything back to me, apply what is safe while flagging the rest, or apply everything and leave only the trace. Whatever the level, the irreversible stays human, and everything applied stays traced and revocable.</p><h3>On your machine</h3><p>Cortex runs locally. Semantic search relies on a model that lives on your Mac. No remote telemetry.</p><p>For its own processing, Cortex needs an engine. It plugs into a subscription you already pay for (Claude Pro or Max, ChatGPT, <a target="_blank" rel="noopener noreferrer nofollow ugc" class="break-all link-break text-gray-500 hover:underline" href="http://Z.AI">Z.AI</a>, Kimi) or into a pay as you go API key. Ollama works too, if you would rather nothing left the machine at all.</p><p>That processing is spread across about twenty workers sorted into three tiers: routine operations, the consolidation core, and genuinely complex analysis. Each tier runs on the connection you assign to it, with a fallback order if that connection drops. So you choose where the effort goes, and what it costs.</p><h3>What it is not</h3><p>Not another chat, not another agent. You do not talk to Cortex: it is a system, not an interlocutor. It is not a memory that learns quietly either, with no way for you to see or correct it. The agent tier, the one that works and acts on the same memory, is called Cortex.deck and plugs in on top.</p><h3>Where it stands</h3><p>Private beta on macOS, priority access on request. Installation takes about ten minutes, most of it downloading: Cortex ships its own tooling and its own model, in its own corner, without touching what you already have installed. First run configuration fits in seven screens, two of which ask for a real decision.</p><p>Two caveats I would rather state than hide. The app is not signed by Apple yet, so macOS shows a warning on first launch. And on the browser extension, capturing ChatGPT, <a target="_blank" rel="noopener noreferrer nofollow ugc" class="break-all link-break text-gray-500 hover:underline" href="http://Claude.ai">Claude.ai</a> and Gemini conversations works, but returning memory into the web page is not live yet.</p><h3>Under the hood</h3><p>A Tauri app, Rust on the system side, React and TypeScript on the interface, with signed automatic updates. The site and API run on Astro as a static build, Fastify and PostgreSQL, deployed in containers behind Caddy.</p>

Cortex · Governed cognition system for AI

<h3><strong>What is arka.norn?</strong></h3><p>arka.norn is a local-first governance and delivery framework for coding agents.</p><p>You keep working in Claude Code or Codex through their normal interfaces. Norn does not replace the coding agent or add another chat. A skill connects the agent to the framework, while Norn keeps the governed project state outside the conversation.</p><p>That state is shared across a local Web cockpit, TUI and CLI.</p><p>The goal is to separate two things that are often mixed together: what the model is free to reason about, and what the system is allowed to accept as project reality.</p><h3>Why I built it</h3><p>I started building products with coding agents about two years ago.</p><p>As the projects became larger, I kept running into the same class of problems.</p><p>An agent could accept a rule, then reinterpret it a few turns later. It could know that a document existed without actually reading it. It could inspect a README and a few directories, then describe an unfamiliar repository with enough confidence to make an incomplete exploration look exhaustive.</p><p>Even simple instructions could produce valid but unwanted interpretations.</p><p>"Fix the bugs" could become disabling what caused the errors.</p><p>"Ship the feature" could become finishing the implementation and calling it shipped.</p><p>"Simplify the homepage" could become removing almost everything except the main CTA and footer.</p><p>The agent had followed the instruction. The problem was that the instruction did not enforce its interpretation, its boundaries, or the evidence required to satisfy it.</p><p>I first tried solving this with more documentation, profiles, procedures, Markdown files and skills.</p><p>Eventually I reached the idea behind Norn:</p><p>A skill carries the method. The system carries the authority.</p><h3>Live framing and durable project state</h3><p>Norn turns the work discussed with the agent into a resumable local plan.</p><p>The human remains responsible for the important boundaries. Framing has two explicit stabilization points: repository grounding, then publication and delivery-route binding.</p><p>Before publication, the framing state stays outside the repository.</p><p>After publication, an immutable fingerprinted plan revision becomes the durable recovery source.</p><p>The chat remains useful for working with the model, but it is no longer the only place where the project state exists.</p><p>This also makes the state provider-portable. A later session, or another supported provider, can recover from the published plan instead of reconstructing the project from a previous conversation.</p><h3><strong>Repository discovery is governed too</strong></h3><p>The same principle applies before code is changed.</p><p>For an unfamiliar repository, Norn can prepare a read-only Project audit.</p><p>The human first selects the decision they are trying to support, the scope and the domains to inspect. Norn prepares the audit plan, and collection starts only after the human reviews and approves it.</p><p>The audit can collect evidence and findings, but it cannot modify Project files or control Agents.</p><p>The model therefore does not get to decide by itself that reading a README and a few directories is sufficient discovery.</p><p>Product documents also retain their provenance, dependencies and history, so legacy documentation can remain accessible without silently becoming the current source of truth again.</p><h3>Web, TUI and CLI</h3><p>Norn exposes the same durable state through three surfaces.</p><p>The Web UI is a local control cockpit for framing, Features, documents, decisions, audits, registered Agents and orchestration state.</p><p>The TUI provides an interactive terminal cockpit.</p><p>The CLI provides exact commands and stable machine-readable output for automation.</p><p>The Web interface is not another AI chat. The agent remains in the provider you already use.</p><h3>Interactive work and automatic orchestration</h3><p>There is an important distinction between the two modes.</p><p>When working interactively in Claude Code or Codex, Norn governs project state but does not pretend to sandbox the provider itself. If Claude Code gives the agent Bash, Bash is still available.</p><p>Automatic orchestration is a separate mode, and that execution is sandboxed.</p><p>Mastra orchestrates the task DAG. Tasks use isolated Git worktrees and bounded read/write scopes, and their effects pass through the Norn broker.</p><p>Agents propose changes. Norn checks whether those changes stay inside the declared scope and whether the required mechanical evidence exists.</p><p>For example, the model can say that the tests pass. Norn can verify that the configured test recipe actually passed and that the resulting diff stayed inside the allowed write scope.</p><p>When the operation is authorized, Norn creates the commit.</p><p>This is the distinction at the center of the project:</p><p>A rule in a skill influences the model.</p><p>A capability that is absent or refused by the system actually constrains it.</p><p><br><strong>Built with Norn</strong></p><p>The first usable version of arka.norn was formalized quickly, but the approach behind it comes from roughly two years of experimenting with coding agents, discarded architectures and real project failures.</p><p>arka.norn is being built with arka.norn.</p><p><br></p>