Context Keeper
A maintenance layer for the markdown brain your agents read.
Every team that handed an agent a knowledge base now runs a second codebase that nobody lints. Documentation folders, rule files, memory files, prompt libraries. They grow every day, they are written by many hands and increasingly by models, and nothing in the toolchain has an opinion about whether a line is still true.
A stale line in a document confuses a colleague. A stale line in an agent instruction makes a machine act on it quietly, many times, without anybody noticing. Context Keeper is the boring layer that went missing while everybody built better places to write things down: not another home for your notes, but something that keeps what is already written honest.
The summary I wrote when I first noted this idea down:
Context Keeper is a check that runs over a folder of markdown files, the kind that people and language models now read as instructions, and says which lines have become false, which contradict each other, which nothing links to any more and which are duplicates of a better line somewhere else. Nothing is deleted automatically. A person answers one question at a time and the answer is kept with a date.
How it would work
It runs in the build, on every commit, and again on a schedule. It reads the files as units rather than as documents, so a single rule line is the thing it can talk about, and it produces one short review rather than a dashboard.
The review has four kinds of finding. Lines that contradict each other, lines that outlived their own review date, orphans that nothing points at any more, and near duplicates where one version is better than the other. Alongside them sit the dead references: files, settings and flags that no longer exist.
Nothing is removed by the machine. Each finding becomes one question with a default answer, and the answer is recorded with a date and a name. That record, not the check, is the part that becomes valuable, because after a year it is the only written account of what the team decided was true and when.
Who it is for
A software team that has already put instruction files in front of a model, so the repository holds rule files, a documentation folder or a prompt library that changed more than once last month. Somebody on that team can already name a line that was wrong for weeks and the thing it caused.
The better paying customer arrives later: the company where a wrong line carries a regulatory or safety cost, and where somebody has to be able to show what was believed to be true on a given date and who signed it off. That customer will pay for the decision record alone and will insist on running the whole thing inside their own network.
The person this is not for is the individual note taker. The problem is real for them and the willingness to pay somebody to prune their own thinking is close to zero, which five years of beautiful note tools have demonstrated fairly thoroughly.
Why now
Hand written instruction text became load bearing in about eighteen months. The volume per team went from one page to several hundred, and the cost of a wrong line went from a confused colleague to an automated action taken at scale.
Every vendor in the category sells writing, storing or finding. None of them sells maintenance, and the reason is that the hard part cannot be automated away: deciding which of two conflicting lines is now false needs a judgement that lives in a person. The product is not that judgement. It is the loop that collects it in seconds instead of an afternoon.
FAQ
Questions people usually ask.
- How is this different from the memory products everybody is shipping?
- Those are built to add and retrieve. This one is built to retire. The unanswered question in that whole category is not what to remember, it is which of two lines is now false, and that needs a person in the loop rather than a cleverer model.
- Could a model not just do this on its own?
- It can find the candidates and it cannot be the judge. A model that deletes what it believes to be outdated will eventually remove something true and confident, and the first time it does that you have lost the customer. The model proposes, a person decides in seconds, and the decision is written down.
- Does this only matter if you run AI agents?
- No, but agents are what made it urgent. A wrong line in a handbook waits politely for somebody to read it. A wrong line in an agent instruction is acted on the same day, repeatedly, by something that does not pause to wonder whether it still holds.
- Why markdown rather than a proper system?
- Because that is where this material actually lives: plain files in a repository, versioned next to the code, editable by anybody. The format is not the problem. The absence of anybody whose job is maintenance is the problem.
- Have you built it?
- I run a version of it on my own knowledge base every day, which is why I trust the design, and it is also why I am not building the company. The dossier says plainly what is running, what it taught me, and which parts of it do not transfer to anybody else.
This idea is for sale
I am not going to build this one.
I have more ideas than years. Leaving this one on a shelf helps nobody, so you can have the whole working file, or you can have the idea itself. Both routes are below, and both are priced so that somebody who is actually going to do the work can afford to start.
Shared licence
The dossier
€99
Everything I worked out about Context Keeper, as a printable PDF you own for life. Written for somebody who wants to start on Monday, not for somebody who wants to be entertained.
- ·The four layers in the order I would build them, and the reason each one has to stay separate from the next
- ·The six rules that decide whether a maintenance layer works at all, every one of them learned by getting it wrong in a system that was already running
- ·The false alarm problem written out properly, since it is the thing that quietly kills tools in this category
- ·Per repository pricing with a worked margin model you can change, and why per seat and per token are both the wrong shape here
- ·The open source wedge: which half to give away, which half to charge for, and why that way round rather than the obvious one
- ·The first ninety days, which start with twenty unsolicited audits before a line of product code is written
A shared licence means you may use and build on everything in the file. It is not exclusivity, other people can buy the same dossier, and it transfers no name, domain or code. Immediate download, so your right of withdrawal lapses on delivery. See the withdrawal terms.
If you would rather have this idea to yourself than share the dossier, say so and we can talk. Be aware that for this one there is no domain, no prototype and no code waiting for you. There is the thinking, and an agreement that I take the idea off this page. Write to me.
Join the early access list
Get a note if this ever moves, or if somebody buys the idea and builds it properly.