The Memory You Own
I caught myself explaining the same thing for the fourth time last month. Not to a person - to an AI.
I was starting a new Claude session to work through a pricing question for sheepCRM. Within the first few exchanges I was re-establishing context I had already provided, in detail, to a different session two days earlier. Who our customers are. How our pricing model works. Why we made the decisions we made about packaging. The AI was patient and competent. It asked good clarifying questions. And I realised I was not annoyed at the AI. I was annoyed at myself, because I had been feeding the same context into three different tools for weeks, and none of them could see what the others knew.
That is the moment this piece is about.
🧠 The second brain, version two
The "second brain" idea has been around for years. Tiago Forte popularised it. The tools are familiar - Notion, Obsidian, Roam, Apple Notes if you are honest about it. The premise is sound: get things out of your head, structure them, and your future self will thank you.
The trouble is that version one of the second brain was built for a human reader. You wrote notes so that you could find them later. The folder structure, the tags, the backlinks - all of it optimised for a person with a mouse, scrolling through pages, making connections in their own head.
That was fine when your tools were passive. It is not fine when your tools are active participants in your work. The AI that helps you write code, draft a proposal, triage a support ticket - it does not browse your Obsidian vault. It does not read your Notion database. It starts every session from zero, and you spend the first five minutes re-teaching it things you have already thought through.
🔒 The renting trap
Here is where it gets interesting, and slightly uncomfortable.
Every major AI platform has noticed this problem, and every one of them has shipped a solution. ChatGPT has memory. Claude has project knowledge and memory. Copilot has its context layer. Each of these works, more or less. Each of them remembers things you have told it across sessions. Each of them makes the next conversation a little smoother than the last.
And each of them locks you in.
Not in the way we usually talk about lock-in. Nobody is holding your files hostage. There is no export fee. The lock-in is subtler than that. What you have built up, conversation by conversation, is a model of how you think - your preferences, your context, your decision history, the shape of your particular problems. That accumulated understanding does not export. You cannot take your ChatGPT memory and hand it to Claude. You cannot take your Claude project knowledge and hand it to the next tool that comes along.
Your switching cost is not your subscription. It is your accumulated context.
I have started calling this comprehension lock-in, and I think it is the most insidious form of platform dependency we have seen yet, because you do not notice you are building it. Every conversation that goes well is another reason to stay. Every session where the AI already knows your background is a small withdrawal from a bank you cannot transfer.
🏗️ What owning looks like
A few weeks ago I built something to fix this for myself. I called it Cairn - a name borrowed from the stone markers hikers leave on mountain trails. A cairn says: someone was here before you, and this is what they found worth marking.
The technology is almost disappointingly mundane. It is a Postgres database. Each entry has a type - a thought, a decision, a source, a person, an organisation, a project. Each entry can link to other entries through typed edges: this thought was informed by that source, this counterpoint challenges that decision, this project supersedes that one. The whole thing is exposed via MCP, so any AI tool I use can read it and write to it.
It took a weekend. The schema is not complicated. The edges are just rows in a join table. The first version had no vector search, no RAG pipeline, no embeddings - just full-text search over a Postgres table, and that was enough. Semantic search came later, sitting alongside keyword search rather than replacing it. None of the usual AI-memory furniture was ever mandatory, and that is the point. The architectural decision is not which database to use. It is whether your knowledge lives in infrastructure you own or in a feature someone else controls.
Now when I start a Claude session, it reads from Cairn. When I finish, it writes back. When a different tool - a scheduled agent, a research assistant, a triage bot - needs context about a decision I made last week, it queries the same store. The context is mine. The protocol is open. The AI is interchangeable.
Hundreds of entries in April alone. Not because I sat down to fill a database, but because the act of working with AI naturally produces things worth keeping - decisions, analyses, research findings, the context behind a customer conversation - and Cairn gives them somewhere to live that outlasts the session that produced them.
⚡ The contradiction advantage
Most knowledge systems try to resolve contradictions. You write two things that disagree, and the system either flags one as outdated, merges them into a synthesis, or quietly lets the newer entry shadow the older one. The instinct is tidy. Tidy is wrong.
The most valuable things in my knowledge store are the tensions. A decision I made in February that I am no longer sure about. A counterargument to my own pricing strategy that I have not yet answered. Two assessments of the same customer situation that reach different conclusions because they were written on different days with different information.
Cairn preserves these as first-class objects. A counterpoint stone has a "challenges" edge pointing at the thing it contradicts. The contradiction does not get resolved. It sits there, visible to every AI that queries the store, and every session that touches the topic has to reckon with both sides.
This is not a design choice born from philosophy. It is a design choice born from watching what happens when you smooth contradictions too early. You lose the signal. The tension between two competing views is often the most important thing you know about a topic, and the moment you flatten it into a consensus, you have thrown away the piece that was actually doing the thinking.
🗺️ The compounding curve
I wrote in The Sum of All Tokens that the cost of software is token spend plus human insight, and that the insight is the scarce half. This is the piece that follows from that.
If insight is scarce, the question becomes: where does it accumulate? If it accumulates inside a platform you do not control, you are renting your own judgment back from someone else. If it accumulates in a database you own, exposed via an open protocol, readable by any AI you choose to point at it - then it compounds.
The hundredth entry in Cairn is not a hundred times more useful than the first. But the graph of connections between entries - the web of "informed by" and "challenges" and "made progress on" edges - that graph is where the compounding happens. A new session does not start from zero. It starts from the accumulated understanding of every session before it, and it can see the contradictions as well as the agreements.
This is The Long Obedience applied to knowledge. Not scattered notes across six platforms, each forgotten the moment you close the tab. Sustained accumulation in one direction. The same direction. For long enough that the graph starts doing work you did not plan for.
The technology is a weekend. The discipline is longer. But the architectural decision - own your memory, do not rent it - is the one that matters, and it is available to anyone willing to make it.
James Webster is the founder of sheepCRM and director of Croftsware. This piece extends The Sum of All Tokens and The Long Obedience.