I had not announced it here yet: I am writing a book about Context.

The cover of the book being written.

Context is the method I am building to let a human and an AI work on a mission over the long term, without starting over at every conversation, while holding on to:

• the sources;
• the knowledge produced;
• the decisions;
• the state of the work and whatever will make it possible to pick it back up later.

The book is now nearly finished.

At that point I could have carried on rereading it, fixing a few sentences, filling out the examples and getting ready to publish.

Instead, I asked myself a fairly uncomfortable question:

what have I really invented in there?

Not "invented" in the sense of having to claim every idea.

The question was much simpler:

am I presenting as a new method something that already exists and that I have simply reinvented without knowing it?

This is first and foremost a matter of the scientific habit of doubt.

When you have been building something for months, it is very easy to look, without noticing, for everything that confirms your idea is good, different or original.

I tried to do the opposite.

To look methodically for everything that could pull that conviction apart.

In everything I present as the Context method:

• what already existed?
• what did I borrow knowingly?
• what did I adapt?
• what did I rediscover without knowing it?
• what had already been put together elsewhere?
• and what, if anything, is specific to the way I brought it all together?

At the start I thought I was simply doing a prior art search.

In practice, what I mostly learnt was how to take my own certainties apart, one by one.

I already knew I had not started from nothing

Some influences were obvious.

The bounded context from Domain-Driven Design is an acknowledged inspiration.

Atomic notes were always going to lead me to Zettelkasten.

PARA and Second Brain shaped the way I think about knowledge spaces.

On the more technical side, I had also knowingly picked up ideas from Git — separation, copying and merging in particular — as well as the Observer pattern for certain relations between Contexts.

So the question was not:

"Are all my ideas new?"

I already knew they were not.

What interested me was the assembly.

A Context, in the book, is a space bounded by a mission. It holds sources, knowledge produced from those sources, a current state, a history, decisions, contradictions, actions, deliverables. It has to be interruptible and then resumable. It has to build up knowledge as the work goes along. And several Contexts have to be able to exchange without losing their autonomy.

So I said to myself: the building blocks surely come from elsewhere.

But maybe the way they combine is distinctive.

That was the hypothesis I set out to knock down.

First mistake: searching with my own words

The first problem shows up straight away when you look for the prior art behind a method.

Nobody was obliged to use your vocabulary.

Searching for "Context" + AI + memory tells you almost nothing about whether a similar mechanism exists.

A medical team may have been doing something close to your handoff for fifty years, without ever using that word.

An archival system may have formalised provenance, source and interpretation without ever mentioning knowledge management.

A piece of software from the 1990s may have solved a copying and dependency problem you think you ran into with AI agents in 2026.

So I had to stop searching for Context.

I started by taking the book apart.

With ChatGPT, we pulled out 24 concepts and 12 structuring relations.

The idea was to turn the book into objects simple enough to ask questions about.

Not just:

Has anyone already imagined a persistent workspace?

But also:

Has anyone already separated the raw source from the knowledge derived from it?

Does any system distinguish the present from the historical log?

Is there a tradition where working on one problem enriches a base that can be reused on the next ones?

What happens when a piece of knowledge is copied from one autonomous space into another?

That changes the search completely.

First surprise: nearly every building block was already there

Once the system was in pieces, the precedents started showing up everywhere.

A few examples:

• provenance obviously had a whole literature behind it, right through to standards like W3C PROV; W3C — PROV-O: The PROV Ontology
• Event Sourcing offers a way to rebuild a current state from the history of events;
• Architecture Decision Records keep a decision, its context and its reasoning, including once it has been superseded; Michael Nygard — Documenting Architecture Decisions
• Truth Maintenance Systems were already working in the 1980s with several sets of assumptions that might well be incompatible. Johan de Kleer — An assumption-based TMS

None of this really surprised me at that stage.

A complex method is rarely made of mechanisms that appeared out of nowhere.

The first conclusion felt comfortable enough:

the building blocks are not new, but the assembly might be.

It was a reasonable conclusion.

It did not hold for long.

Second pass: search for the function, not the concept

We changed method.

Instead of searching for the book's concepts, we started searching for the problems they answer.

• How do you pick work back up after an interruption?
• How do you hand a case over well enough for someone else to carry on?
• How do you turn thousands of scraps of information into evidence, and then into hypotheses?
• How do you hold several competing explanations without artificially picking one?
• How does a workspace hold its operational state and reusable knowledge at the same time?

This time, the search left software very quickly.

It took us towards:

• intelligence work;
• law;
• medicine;
• incident management;
• cognitive science;
• archival science;
• scientific research.

The sensemaking model of Peter Pirolli and Stuart Card, for instance, describes a process in which the analyst starts from external sources, picks out the relevant information, builds up evidence, constructs schemas and hypotheses, then produces the result of the analysis. Pirolli & Card — The sensemaking process and leverage points for analyst technology

This is obviously not Context.

But the closeness of the mechanism was interesting.

Same with Jakob Bardram's Activity-Based Computing.

In 2005, Bardram describes an infrastructure where the central unit is no longer the application but the human activity, together with its associated computing tasks and resources. The architecture is designed in particular to support mobility, collaboration and continuity of activity across environments. Aarhus University — Activity-Based Computing: Support for Mobility and Collaboration

In other words, the idea that you need to keep something richer than a document if work is to carry on had obviously not waited for AI agents.

Medicine, command centres and incident management have their own handoff mechanisms too.

The ground was already starting to shrink.

Picking work back up was not distinctive.

Nor was the handoff.

Even the chain:

source → interpretation → reasoning → output

had very solid precedents.

So, a new hypothesis:

maybe what is distinctive sits at the junction between the live case and the body of knowledge.

In other words: working on a problem while building something that will outlive the problem.

That hypothesis held for even less time.

The detour that changed the search most: leaving English

Until then, even when we changed field, we were still largely shut inside a Western, English-language corpus.

So we ran searches in Japanese, Chinese, Russian and Arabic.

That is where the search got genuinely interesting.

In Japan, I came across KJ法, the KJ method, from Jirō Kawakita.

Kawakita describes a bottom-up method in which you start from field data, write it onto cards you can move around freely, then gradually bring them together and combine them by meaning so that a structure emerges. J-STAGE — Emergence and Essence of the KJ Method, an interview with Jirō Kawakita

And there I had a rather amusing moment.

A few months earlier, knowing nothing about KJ法, I had myself started working with what I call pillar cards.

I wrote about them here recently: Pillar cards: keeping your memory alive so you can keep thinking

It is not the same method.

KJ法 is used, among other things, to start from scattered raw input and let a structure emerge gradually through grouping and recombination.

My pillar cards are there instead to keep a handful of structuring ideas alive, to revisit them and to keep testing them against what I read and try out.

But there is still a family resemblance, and it interested me a great deal:

• getting units of thought out of your head;
• making them movable;
• keeping them as relatively autonomous objects;
• bringing them together and recombining them;
• using those juxtapositions to keep thinking.

And above all: I discovered KJ法 long after putting that practice in place.

That is exactly the kind of thing I wanted this search to turn up.

Not so I could say: "I had the same idea."

Quite the opposite.

So I could see that you can honestly build something on your own and then find out that another tradition has been exploring a neighbouring mechanism for decades.

We also found Tadao Umesao, who describes as early as 1969 a system of independent cards meant to be recombined, where the new knowledge produced by that recombination can itself go back into the system.

In other words:

existing body of knowledge → recombination → new knowledge → richer body of knowledge

That was already getting close.

Then we arrived in China.

And there I discovered the Hall for Workshop of Meta-Synthetic Engineering, or HWMSE.

HWMSE: where the hypothesis really started to fall

Meta-Synthesis is developed around Qian Xuesen in the early 1990s to deal with complex systems.

HWMSE becomes the environment meant to put it into practice.

Its architecture explicitly combines:

• human experts;
• machine systems;
• knowledge systems.

The aim is not only to use existing knowledge.

The authors also talk about knowledge creation.

They distinguish a general knowledge system from a knowledge system specific to the object being studied, and explain that the second has to be able to build up continuously through its own use. Yu, Zhou & Feng — Man-machine collaborated knowledge creation in HWMSE

That already looked a great deal like what I was trying to set out with Context.

And it was not only a conceptual diagram.

The HWMSE work led to prototypes and experimental applications, around macroeconomic decision-making in particular. JAIST — Knowledge System Construction and Knowledge Creation in HWMSE

The comfortable distinction:

case management systems handle the work, knowledge systems handle the body of knowledge, Context brings the two together

no longer really held.

Others had already built systems that brought together a substantial part of both.

I had to go down another level.

Third pass: look for the architectural rivals

We stopped asking:

"Which concept looks like Context?"

Instead, we set out a few signatures:

• does the system handle work situated around a problem?
• does it turn information into knowledge?
• can that knowledge serve other cases?
• does it distinguish several kinds of memory?
• can several spaces cooperate while staying autonomous?

That search turned up several systems far closer than expected.

Experience Factory

The Experience Factory, developed around Victor Basili's work and the experience of NASA's Software Engineering Laboratory, explicitly separates two functions.

On one side, the project organisation delivers its project.

On the other, the Experience Factory:

• collects the experience produced;
• analyses it;
• synthesises it;
• turns it into reusable units;
• makes it available again to future projects.

NASA Goddard's Software Engineering Laboratory was already described in 1992 as an operational example of an Experience Factory, after fifteen years of running and several hundred studies. NASA — The Software Engineering Laboratory: An operational software experience factory

Basili and Caldiera then explicitly describe the Experience Factory as an organisation separate from the projects, whose job is to capitalise on experience and reuse it. NASA — The Experience Factory: Strategy and Practice

That struck directly at another idea I thought mattered:

the deliverable is not the asset; the project's work has to enrich something that will outlive it.

The Experience Factory was already saying that.

Knowledge-Centered Service

Then Knowledge-Centered Service, or KCS, offered another version of the same mechanism.

In KCS, knowledge is not meant to be produced in a process separate from the work.

The solve loop consists, among other things, of:

• searching for existing knowledge;
• reusing it;
• improving it as it gets used;
• capturing new knowledge where needed.

The Consortium for Service Innovation sums the idea up in a particularly clear phrase: reuse is review. Knowledge is checked and improved in the flow of the work itself. Consortium for Service Innovation — The Solve Loop

Another reduction.

Even:

making everyday work a machine for producing and improving reusable knowledge

had already been formalised.

Fourth pass: going after the last details

At that stage, piling up more broad analogies was not worth much.

So we picked a few elements that still seemed relatively distinctive:

• the strict separation between raw source and derived knowledge;
• the state / history / synthesis / handoff split;
• the atomic note as a sourced claim that can be supported or contradicted;
• the difference between referencing a piece of information and copying it, then taking on its maintenance.

Even there, the walls came down one after another.

Nanopublications and micropublications already represented small assertions carrying provenance, supporting elements and sometimes challenges.

Federated database systems had long been working on the difference between remote data and a local copy that may drift.

Cognitive architectures distinguished working memory, episodic memory and semantic memory.

Forensic computing practice keeps the original and works on copies in order to preserve the evidence.

Each time, something that still looked distinctive turned out to be a recombination of known mechanisms.

The last real find: ScholOnto and SWAN

Even so, we had to know whether we had reached the end or were simply tired of looking.

So we added a criterion.

A new source no longer counted just because it looked like Context.

It had to change the map.

For instance:

• be older than anything we had already found;
• be markedly closer;
• bring together several mechanisms that had been separate until then;
• or force us to change our conclusion.

That rule let us carry on without falling into the game of forcing any analogy that comes along.

Two systems still passed the test.

ScholOnto

The ScholOnto project, funded from 2000 and active in the early 2000s, set out to build an infrastructure that would let researchers publish claims, connect them to the literature and challenge them.

Its ClaiMaker system let you build and navigate those networks of claims and arguments. Knowledge Media Institute — ScholOnto / ClaiMaker

SWAN

Then SWAN, in biomedical research, went further still.

SWAN explicitly distinguished external scientific resources from the discourse elements produced inside the system.

A Research Statement could be derived from an article or another resource, keep its provenance, and be linked to other statements through relations such as supports, refutes or alternativeTo.

The system also kept earlier versions when a statement changed. Ciccarese et al. — The SWAN Scientific Discourse Ontology

At that point, the closeness to my atomic notes was hard to ignore:

• a distinct source;
• interpretation;
• claim;
• provenance;
• support;
• contradiction;
• reuse;
• change over time.

And here again, this was not only a theoretical ontology: the authors describe a Workbench and a Browser used to build and navigate that knowledge base. PubMed — The SWAN biomedical discourse ontology

Most of the pieces were already there.

And then the results stopped changing the conclusion

This is probably the most important part of the method.

At some point, you have to stop.

Not because you have proved no other prior art exists.

That is impossible.

But because new searches start returning the same families over and over.

One more paper improves a date.

Another source adds an example.

A system uses different vocabulary but covers a mechanism already identified.

Or worse: you find yourself spending ten lines explaining why something might vaguely be comparable.

At that point, there is a risk.

You are no longer putting your hypothesis to the test.

You are starting to manufacture the resemblance you wanted to find.

So we took the view that the search had reached reasonable saturation.

So what is original?

The conclusion is not the one I imagined at the start.

I no longer think it is very useful to look for "the new idea" inside Context.

Nearly all its important building blocks have precedents.

So do many of the relations between those building blocks.

And several earlier systems already assemble large chunks of them.

For instance:

• HWMSE brings together humans, machines, problem-solving and building up knowledge;
• Experience Factory separates the project from the organisational mechanism that turns its experience into an asset;
• KCS builds the creation and improvement of knowledge directly into everyday work;
• Project Memory works on the history and context of the project;
• ScholOnto and SWAN give structure to claims, sources, provenance and contradictions;
• federated systems work on autonomy, import, copying and dependency.

That does not mean Context brings nothing.

It means its contribution lies elsewhere.

Today I would put it like this:

Context is an explicit selection, combination and normalisation of mechanisms from several traditions, brought together into a working method for use with an AI over the long term.

The system has, among other things:

• its vocabulary;
• its objects;
• its constraints;
• its procedures;
• its particular way of separating source, knowledge, state, history, synthesis, decision, contradiction, handoff and deliverable;
• its rules for moving information between several autonomous Contexts.

I did not find a historical system that has exactly all of that grammar.

But nearly every component of that grammar has a history before Context.

And in the end, that is an answer that suits me much better.

The scientific habit of doubt

At the start, the question looked something like this:

Is my idea really original?

Looking back, that is probably not the best question.

What the prior art search mostly did was let me understand what my idea was made of.

Which intuitions came from traditions I knew.

Which ones existed in fields I would never have thought to look at.

Which pieces had already been put together.

And which part of my work was mostly about translating, simplifying, recombining and making usable a set of scattered mechanisms.

More than anything, it is a good illustration of what I mean by the scientific habit of doubt.

I had an implicit hypothesis:

maybe this part of Context is new.

The right way to test it was not to look for reasons to believe it.

It was to look for the best possible counter-example.

And then, when we found one, to change the hypothesis.

With every pass, the ground I could reasonably present as specific got smaller:

• first the building blocks;
• then certain assemblies;
• then the idea of making work itself a way of building up knowledge;
• then part of the information grammar itself.

Paradoxically, the claim became clearer.

I would rather be able to say:

here is what I borrow, here is what I adapt, here is what I combine, and here is the shape I give it

than tell people an idea is new simply because I had never met its ancestors.

The method I would use again

If I had to run a search like this on another idea, I would not start the way I did this time.

I would go straight to:

• breaking the idea down into components and relations;
• searching for the mechanisms, not its vocabulary;
• comparing components, relations, assemblies and purposes separately;
• leaving the original field quickly;
• searching in several languages;
• looking at systems that were actually built, not only theories;
• working back through the bibliography of the best sources, then looking at the work that cites them;
• deliberately looking for the counter-examples that could bring the current conclusion down;
• defining what a new find has to change to count as important;
• stopping once new sources only bring variants of precedents already on the map.

Working back through the references of a good source and then following the publications that cited it is sometimes called snowballing.

The word matters less than the mechanism.

A good source becomes a way into a whole vocabulary and a whole tradition that a keyword search might never have found.

The last point — knowing when to stop — is probably the hardest.

A search can always carry on.

There will always be:

• another paper;
• another patent;
• another field;
• another language;
• another keyword.

The question is not whether everything has been found.

The question is:

does carrying on still change the answer?

The limits

This method obviously does not prove that no prior art exists.

It depends on:

• the search engines;
• the languages you think of;
• the documents that have been digitised;
• the publications still available;
• how they are indexed;
• the vocabulary their authors chose;
• and our own ability to recognise the same mechanism behind two different vocabularies.

It also depends on the judgement used to decide that two mechanisms really are comparable.

A resemblance is not an identity.

A technique used in a nuclear power station is not the direct ancestor of a memory system for AI simply because it has a handoff.

In the same way, finding out after the fact that a Japanese method uses cards does not make my pillar cards KJ法.

You have to be able to say both:

this is close enough to be interesting

and:

this is not the same thing.

It is also why finding nothing has to stay finding nothing.

Not turn into:

nobody had ever done it.

What I take from it

I started this search trying to find out what I had invented.

I am finishing it with a slightly different question:

what have I built out of everything that already existed?

The answer strikes me as more interesting.

Context is probably not interesting because it discovered:

• provenance;
• atomic notes;
• the handoff;
• organisational memory;
• bounded spaces;
• building up knowledge.

All of that existed.

It becomes interesting if the assembly makes those mechanisms simple and coherent enough for a human and an AI to work together for weeks or months without losing what they learn.

So I did not come out of this search with virgin ground to plant a flag in.

I came out with a map.

And to understand what you are building, a good map is probably worth a great deal more than a flag.