Personal Knowledge Management for Conversations

Twenty years of tooling exists to move text from where you found it to where you keep it. Talk arrives in the wrong format for any of it.

Anything written already has plumbing. A clipper for the article. Highlight sync from the e-reader. A share sheet on the phone. Copy and paste, so ordinary that nobody counts it as a feature.

All of them work for the same boring reason: what you are moving is already text.

What someone said to you is not text yet, so the road stops there. The client explaining what they actually want, the colleague telling you why the plan will not work — that material has nowhere to go.

The familiar complaint about note systems is that things go in and never come out again, which is how the vault fills up and stays closed. This sits one step earlier. The material never gets in.

A whole industry turns meetings into notes. Every tool in it works by joining the call, which is the one thing an in-person conversation does not have.

If it happened in a Zoom window you are well served. A bot joins, the transcript lands, the summary is in your notes tool before you close the tab.

All of it hangs on one thing: a link to join. There is no link for a car, a corridor, or the ten minutes in the car park afterwards where the real objection finally comes out.

And if your work is the kind where the deciding conversation happens in a car park, that entire industry has nothing for you.

In a room you choose between being present and having a record. That is only the first of four things working against you.

You cannot be in it and record it

On a video call you can type through the whole thing and nobody minds. Across a table, reaching for a laptop changes what the other person tells you. So you pick one: be in the conversation, or keep it. Almost everyone picks being in it, and loses the content.

It has no edges

A scheduled call arrives pre-labelled. Title, time, attendees, a start and an end that somebody else defined.

A corridor conversation has none of that. Even when you do record it, it lands as an untitled forty minutes, and an untitled forty minutes is a file you will never open again. Six weeks later what you have is "someone told me the integration was the real blocker", and no way back to the file.

The useful part is often after the end

You leave the meeting, walk to the car, and somewhere between the lift and the street it resolves into one sentence. That sentence lasts about ninety seconds in your head. Nothing has ever caught it.

You do not know which one it was

The conversation that turns out to matter is usually the five minutes before the scheduled one started. Any system that asks you to decide up front which conversations to keep is asking for a prediction nobody can make.

Most of it comes down to making the start cheap. If beginning a session costs one tap, you can afford to be wrong about which conversations mattered.

If you are sitting across a table from someone, the only thing that can be asked of you is a tap. A Watch complication does it, or your phone lying face down between you. No laptop opens, nothing joins anything, and it never begins on its own — you start it, and you stop it.

Then you can stop naming things. The day comes back to you as Activity, Highlights, an Action Board, Reflection and Communication, with the full transcript underneath, and six weeks later you can go looking the way you actually remember it. Search runs on meaning rather than exact words, so "the call where the pricing objection came up" finds the session even though nobody ever typed those words.

The walk to the car is the part worth leaving it running for. When the sentence finally arrives, say "Pastok" and then say the sentence, and it comes back at the top of the day's Notepad rather than buried at minute forty-three. It only does that when you say the word.

Which means you can start on a hunch and be wrong most of the time. What you keep is text, not a growing pile of recordings — audio is deleted after processing, within seven days by default, while the transcript and everything built from it stay until you delete them. A session that turns out to be nothing costs you the tap.

If you are weighing this against a dedicated recorder or a wearable, what to look for in one goes through that comparison properly.

An archive is what you get by default. Turning it into something you use takes one written instruction about what to lift out, running without you.

Capture on its own just moves the pile. A year of transcripts sits there the way a shoebox of receipts sits there, and you keep meaning to sort it.

The step that changes it is deciding what should be lifted out, once, in writing. In Pastok that is a Skill: an instruction in plain language that runs on demand, on a schedule, or when a situation you described shows up. Each one reads its own window of history and writes into your daily report.

Every Friday, look at the last two weeks. List only the things I said I would do that I have not mentioned since, and who I said them to. Do not summarise the meetings.

That one is a promise-tracker. Another that people write is "tell me which objection has come up three times now, and quote the exact words each person used", which catches a pattern nobody spots unaided.

Long-term Memory carries the recurring people and threads across days, so a later report leads with what is still open rather than starting from zero.

And if you already work inside an AI tool, the MCP server is live and free on every account including ones that have never paid, reaching the last seven calendar days. Claude Code, Codex and Cursor can read your transcripts and reports inside that window, read-only, with the setup steps for each client written out. Anything older than the week you work with inside Pastok.

One thing to settle before any of this: recording other people means asking them first, and where you live that may also be the law.