The work needed its context explained again every time.
An AI could do a useful share of my product work. It could draft the spec, weigh the tradeoff, or summarize the thread, but only after I had explained my world to it again: which decision we had already made, which one we had reversed, and what the last customer call changed. Every session started cold. Sometimes it answered confidently from a fact that had been true a month earlier but was no longer current. The help was real, but the effort required to reach it was often not worth it.
I thought retrieval would solve that problem. If the model could reach into everything I had already written, I would not have to keep repeating myself. So I built a retrieval product.
I built a retrieval product. It worked.
Before Phronesis, I built Product Shadow, a hosted context product for product managers. It indexed a team's documents and conversations and pulled relevant passages into an AI chat on demand. It used retrieval-augmented generation — RAG — to search the team's material for passages that matched a question and supply them to the model. It exposed that context through MCP, the open interface AI tools use to reach outside sources. It was multi-tenant, ran in production, and did its job: ask a question, and it put the relevant passages in front of the model.
Using it in my own work made its limits clear.
Then I stopped using it.
I stopped using it because retrieval answered a narrower question than I needed answered. It could find what I had written, but it could not tell me what was still true. Every new conversation made the model reconstruct the current state of my work from whichever passages came back: which statement was current, which decision had superseded another, and what mattered now. Search made the archive easier to reach. The model still had no maintained account of my work. A current passage and an outdated one arrived looking the same; search could not know that a meeting two weeks later had reversed the paragraph it surfaced.
I rebuilt the approach more than once to close that gap. Eventually I saw that the missing work was maintaining a current account of what the documents added up to: what was true, what was noise, what had changed, and what it meant now.
A simpler tool showed what mattered.
Late in 2025 and early in 2026, I spent time with OpenClaw, a file-native agent environment that kept its instructions, memories, daily notes, and standing work in plain files owned by the user. It was simpler than the retrieval service and often more effective for my work.
It had reliability problems. The useful lesson was in its file-native design: durable behavior could come from a legible workspace you own, without depending on a service you query.
The version I preferred was already on my disk.
I was already using that approach in my own work. My workspace in Claude Code — structured Markdown, explicit instructions, a growing memory, and a handful of skills — had become the tool I reached for first. By February 2026 I had written down the conclusion: build the next version around the workspace I was already using.
I had to treat the production architecture as evidence, then move past it.
The missing operation had a name.
A couple of months later, I found a useful name for that work. In April 2026, Andrej Karpathy described using a language model to maintain a knowledge base: incrementally compiling raw material into a structured account, maintaining the index and relationships, checking for contradictions and gaps, and filing new conclusions back in.
I had made that decision in February. Karpathy's description gave the operation a name: compilation. The loop was now clear: capture, compile, retrieve, correct, compile again.
Retrieval finds what was said. Compilation maintains a current, source-linked, revisable view of what is true now and why it matters.
One product, on files you own.
By early May 2026, I had combined these ideas into one product: a workspace on your own disk for personal, product, and business work. Raw material accumulates in one layer — chats, documents, data, meetings, and recurring agent runs. The system compiles that evidence into a current account: what is true, which source supports it, what has been superseded, where the evidence is thin, and what the situation calls for now. Checks can flag a claim with no source or two conclusions that disagree.
As Claude Code improved at recurring and looped work, I needed less of a separate always-on agent runtime and could consolidate around the files. The loops covered only part of that runtime gap. The result was one product, with a door for each domain, built on files I own.
What Phronesis does.
Phronesis uses retrieval to find source material, then maintains a current account of what that evidence means. It shows you what changed so you can correct it, and lets that context travel between models.
Retrieval finds what was said. Phronesis turns that evidence into a current, source-linked account of what is true now, what changed, and why it matters. You can change models while the account and the trail from evidence to conclusion remain legible and yours.