Why not just memory?

Why not just use Claude or ChatGPT memory?

Their memory lives in their vault, runs on their model, and dies when you switch tools. Craft Phronesis is the inverse: the workspace is a folder of Markdown files on your disk, it runs against whatever model you point it at, and it is still yours the day you leave. Their memory makes their product stickier. Phronesis makes your judgment portable. Those are opposite designs, and only one of them survives you changing your mind about which AI you trust next year.

There is a deeper difference than storage. Chat memory remembers what you told it. It does not catch the decision you made in passing, compile scattered captures into current-state you recognize as yours, or bring back the one prior call that rhymes with the one in front of you. Remembering the conversation is not the same as compounding your judgment.

Why not Notion AI, now that it has custom agents and memory?

Notion is constitutionally a blank canvas, and its memory is built on Notion pages in Notion's database. That is platform lock-in by design: the richer your memory gets, the more expensive it is to leave. It is also, as of May 2026, a paid credits add-on, so the memory you build is metered by the platform that owns it. Phronesis ships the opposite trade. Nothing you build is trapped, because there is no Phronesis-side vault to trap it in.

Isn't this just better retrieval? Better RAG?

Before Phronesis, I built a hosted retrieval system — retrieval-augmented generation, or RAG — for product managers. It was multi-tenant, ran in production, and exposed retrieved context to AI tools through MCP, the open interface those tools use to pull in outside context. It found relevant documents and brought them into a model conversation. I stopped using it because the model still had to reconstruct which statement was current, which decision superseded another, and what mattered now.

Phronesis gives retrieval a maintained account to search alongside the source material. That account records what is current, what changed, and where each conclusion came from; you can read and correct it. The workspace holds the evidence and current account. The codex holds the durable principles and aims that help the system understand what bears on the moment. Retrieval can then bring back a smaller, sharper account without reinterpreting the whole archive every time.

The origin story explains how a retrieval product I built and shipped led to that distinction.

What happens when the frontier model gets good enough to do all of this itself?

Part of it already can, and that is deliberate. Extracting the repeatable playbook from interaction logs is close to a commodity now (Brynjolfsson et al., QJE 2025, is the peer-reviewed proof). Phronesis does not bet against that. It stands on the other half: the situated judgment you bring to a conversation that happens once, which thread to pull, which thirty seconds carried the load, compounded into a workspace you own. That half is durable for a principled reason, it lives in what you did in an unrepeatable moment rather than in a recorded pattern. And whatever the next model can do, it does it better against a workspace that already holds your judgment. A stronger model is fuel, not a threat.

What do I own?

Everything, and it is enforced, not promised. Three ownership gates sit in the release path: export (tarball the workspace and it renders in Obsidian and VS Code with zero Phronesis dependencies), model-swap (change one config value to a different provider), and skill-portability (a skill runs under raw Claude Code, Cursor, or Codex CLI with no Phronesis runtime). If any of the three fails, the release does not ship. The moat is yours or it is nothing.

Then why not just do it myself with Obsidian, folders, and an API key?

You can. The file-native version of Phronesis grew from a hand-built workspace that followed the production retrieval product and made the missing work concrete. Phronesis bootstraps and maintains that system for you: a workspace shaped for your domain, a current-state layer you can review and correct, and the skills that keep it useful over time.

Why not just live in a general AI operator environment (OpenClaw, Hermes, and the like)?

Don't leave it. Feed it. If one of those is already where your context accumulates, the move is not a second workspace fighting for the slot. It is a substrate that makes the one you have dramatically better. A general operator environment running on thin context is a capable agent guessing. The same environment pointed at a Phronesis workspace acts from your caught decisions, your stakeholder graph, your validated taste, the judgment you have compounded over months. That is the gap between an agent that sounds right and one that is right because it knows what you concluded and why. Phronesis is portable and model-agnostic by design, so it plugs in under whatever environment you already trust and hands it something worth acting from. You do not switch. You upgrade what you already run.