Digital Sovereignty

An embedding is not a memory

The National Archives holds two surveys of Britain. One was written on parchment in 1086 and can be read today by anyone who can read Latin. The other was published in 1986 and had to be rescued by an emulation project. Organisations are now accumulating what their AI systems learn in the second format, and mostly do not know it.

P J Laszkowicz

On 4 January 2025, OpenAI marked text-embedding-ada-002 as deprecated, with retirement no earlier than the following June.1 For most of the teams affected this was an ordinary piece of operational housekeeping: read the notice, schedule the work, migrate to a newer model that was cheaper and measurably better. The replacement was an improvement on every axis anybody was measuring.

What the notice actually required, though, was not a configuration change. Every organisation that had built a search index on those embeddings had to compute them again from scratch, because a vector produced by one model cannot be reinterpreted by another. There is no conversion utility. The stored numbers are not a compressed form of the text; they are the position of that text in a coordinate system that exists only inside one particular model. Change the model and the coordinates refer to nothing.

Teams that had kept the source text alongside the vectors re-ran the job and absorbed the compute bill. Teams that had not kept the source text discovered that they were holding several million numbers whose meaning had been retired.

I should say plainly that we build governance and agent infrastructure, and that this argument came out of designing our own approach to the problem rather than from observing it at a distance. Some of what follows describes a test we have not yet run at scale, which is a particular kind of interest to declare.

Two surveys of Britain

In 1086 the clerks of William the Conqueror completed a survey of England and parts of Wales, recording the holdings, populations and values of some thirteen thousand settlements. The result is held today by The National Archives at Kew. It is written in abbreviated medieval Latin on parchment, and reading it requires expertise, patience and reasonable light. It requires no machine at all.

In November 1986, to mark the survey's nine-hundredth anniversary, the BBC published a second Domesday. It was a genuinely ambitious piece of work: around a million people contributed, mostly schoolchildren surveying their own localities, and the result was a multimedia portrait of 1980s Britain covering geography, society and daily life. It shipped on two twelve-inch LaserDiscs, read by a BBC Master computer fitted with a specific interface and a specific disc player.2

By 2002 there were serious fears the discs had become unreadable. The machines capable of driving them had become rare, and the drives rarer still. The problem was worse than obsolete hardware. The still images had been stored on the disc as single-frame analogue video, and the computer overlaid its own graphical interface on top of them at display time, which meant the pictures and the software that interpreted them were not separable artefacts. You could not simply read the disc and recover the content; you needed something that behaved like the original system.

The rescue came from the CAMiLEON project, a collaboration between the universities of Leeds and Michigan funded from 1999 to 2002, which extracted the raw data from surviving discs, around 70 gigabytes per side, and built an emulation of the BBC Master and its disc player so the material could be read on ordinary computers. A separate effort tracked down the original one-inch videotape masters.3

The comparison is not that the older technology was better. Parchment burns, and a good deal of it has. The comparison is narrower and more useful than that. The 1086 survey is legible because interpreting it requires only knowledge that is widely held and independently maintained. The 1986 survey became illegible because interpreting it required a machine, and the machine went away on a schedule nobody at the BBC controlled.

Sixteen years, against nine hundred and forty.

What an embedding is, and is not

An embedding is a list of numbers describing where a piece of text sits in a space the model has learned. Similar texts land near each other, which is what makes vector search work: you embed the query, find the nearest stored points, and return what they came from.

The space is a property of the model. Two models trained on different data, or the same data with different objectives, or the same objectives at a different scale, produce different spaces. The number 0.0731 in the fourteenth position of a vector means something in one and nothing in another. This is not a compatibility problem of the sort that a standard could fix, in the way that a document format can be standardised so two programs agree on what a paragraph is. It is closer to two people using the same word for different things, except that neither is available to be asked.

Which means that when a team stores embeddings and treats them as its memory, it has stored a projection of its knowledge onto a coordinate system that belongs to somebody else. The vendor is not obliged to keep that system available, and in the ada-002 case gave twelve months' notice before withdrawing it, which is considerably more courtesy than the industry norm.

The vendors' own migration guidance makes the position clear enough. Qdrant's documentation on changing embedding model states that switching requires re-embedding every vector in a collection, and assumes throughout that the necessary source data is stored in the payloads, because without access to the original content the vectors cannot be regenerated at all.4 That assumption is doing a great deal of work, and it is one many systems do not satisfy.

Four places memory lives

Agent memory in 2026 accumulates in four places, and each is shaped by something replaceable.

Provider-native memory lives with the vendor. Anthropic's memory tool gives a model a directory it can read and write across sessions, and paired with server-side context editing the company reports token savings of around 84 per cent on long-running tasks.5 It is effective and it is Claude's.

Framework memory objects are written against one library's schema. The mature options (Mem0, Zep with its temporal knowledge graph, Letta, Cognee) are capable systems, and every one of them lives inside the agent process and is reached through its own software development kit. This is why none of them is portable across frameworks in practice: changing agent framework changes the memory layer with it, because the memory layer was always a component of the framework rather than a store the framework happened to use.

Vector indices are shaped by the embedding model, as above.

Summaries and distillations carry the idiom of the model that wrote them. A précis produced by one model reflects what that model found salient, and a system that stores the summary and discards the transcript has made an editorial decision it cannot revisit.

The pattern across all four is the same, and it is the Domesday pattern. The artefacts holding an organisation's accumulated knowledge are precisely the artefacts invalidated by replacing the component underneath them. And unlike a LaserDisc, none of them looks like a preservation risk. A directory of vectors looks like data. A memory store with an API looks like infrastructure. Nothing about either announces that it is legible only to a particular machine.

Anthropic exports, OpenAI does not

Something worth noticing has happened commercially, because it suggests at least one vendor has read the situation correctly.

Anthropic has made memory exportable and importable in a documented form, and has published a route for bringing context in from other tools. OpenAI, as of writing, provides no equivalent API for exporting the memories its products accumulate about a person.6 The gap is not a technical one, since an export is not hard to build, and its persistence at one company while a direct competitor closes it is the kind of thing that only stays stable while it is nobody's problem.

That is a small commercial divergence and it points at a larger one. If the accumulated context of an organisation's work is the thing that makes the next piece of work cheaper, then whoever holds it in a form only they can read holds the account. This is the same structure as the argument about gateways, one layer down and considerably harder to reverse, because a routing decision can be changed in an afternoon and two years of accumulated context cannot be reconstructed at any price.

vec2vec

There is research suggesting the embedding problem might be less absolute than it looks. What it actually demonstrates is not what a hopeful reader would want.

In 2025, Jha, Zhang and colleagues published Harnessing the Universal Geometry of Embeddings, describing a method called vec2vec that translates text embeddings from one model's space into another's with no paired data, no access to either encoder, and no predefined matches. The translations reach cosine similarities as high as 0.92 against native embeddings, across model pairs differing in architecture, parameter count and training data. The work rests on what its authors call the Strong Platonic Representation Hypothesis: that encoders, whatever their design, converge on nearly the same underlying representation.7

Read as an engineering result, this looks like a migration path. Read as what it is, it is a security finding, and the authors present it that way. If a stored vector can be translated into another model's space without the original encoder, then a stored vector leaks information about the text it came from to anyone holding the vectors. The industry has spent several years treating embeddings as a form of obfuscation, the reasoning being that a list of floating-point numbers is not personal data in any operational sense, and this line of work erodes that assumption directly.

So the position is worse than it first appeared. Embeddings are not portable enough to be relied on as a store, and not opaque enough to be relied on as a protection. Building on either property is building on something the research is actively removing.

The test

The useful question is not which memory product to buy. It is a diagnostic that any team can run against whatever it already has.

Delete every index, every embedding and every generated summary. What have you lost?

If the answer is nothing, because all of it can be recomputed from material you hold in a form that needs no particular model to interpret, then the knowledge is portable and a change of embedding model is a reindexing job. If the answer is anything at all, then that part of your organisation's memory exists only in a projection, and its lifespan is bounded by the component that produced it.

This sorts artefacts into two kinds. There is the material that is a source of truth: the observation, the decision, the correction a person made, the text somebody wrote, along with where it came from, when it was true and whether anybody has since confirmed it. And there is everything computed from that material to make it fast to search. The first kind should be stored in a form a person can read without tooling. The second kind should be treated as a cache, however expensive it was to build.

Stated that way it sounds obvious. It is not what most systems do, because the fast thing is the thing you interact with daily and the slow thing feels like an archive nobody visits. The Domesday team made the same trade for the same reason: the interactive experience was the point, and the fact that it could not be separated from the machine that rendered it was invisible until the machines stopped being made.

Two further properties follow. Both are cheap to build in and expensive to retrofit. A stored embedding should carry the identity of the model that produced it and the text it was computed from, so that a change of model is a job rather than a loss. And retrieval should have at least one path that does not depend on a model at all: following links a person curated, or plain term search. That way a change of model degrades quality instead of removing the ability to find anything.

Four attempts at the same standard

The gap has been noticed, and the response has the shape these things usually have at this stage.

There is a format, the Modeled Information Format, at version 1.0.0 since June 2026 under an MIT licence, which keeps markdown as the canonical representation with a machine-readable projection regenerated from it, models validity across two time dimensions, and stores an embedding alongside the model and source text needed to recompute it.8 There is a W3C community group on agent memory interoperability, chartered in June 2026 with seventeen participants, working on interoperability profiles and a regulatory crosswalk.9 There is an Internet-Draft submitted to the IETF's Independent Submission stream, not an IETF standard and explicitly not endorsed as one, specifying encrypted, revocably shareable memory cells whose lifecycle is anchored on a public blockchain.10 For the adjacent problem of what an organisation's data means, a vendor consortium format reached the Apache incubator in June 2026 with more than thirty organisations behind it.11

Two of these efforts share an originator. Most have very few people. None has meaningful adoption yet.

I have watched a version of this from close range. Before the current crop of knowledge formats existed we were building on JSON-LD, on the straightforward reasoning that a document carrying its own semantics in a widely implemented serialisation would outlive whatever happened to be reading it that year. Some of that was work on Solid, where the entire premise is that a person's data should remain legible independently of the application that produced it.

The formats moved, as formats do. What survived the moves was not any of the tooling. It was that the documents were plain text, they carried their own meaning, and when the thing reading them was replaced the documents did not have to be. Migration meant writing a new reader. It never once meant losing the material, and at the time that felt like an unremarkable property rather than the whole point.

The history of metadata standards suggests how this resolves, and it is not by technical merit. Dublin Core, RSS against Atom, microformats against RDFa against schema.org: in each case what settled the question was which format the large consumers actually read, and the survivors were consistently the ones a person could inspect without special tooling. The blockchain-anchored proposal is the most technically ambitious of the four and the least likely to be adopted by anyone operating in a disconnected environment, which is a substantial share of the organisations that care most about this problem.

If the pattern holds, the format that wins here will be plain text with structured headers, and it will win for the same reason parchment outlasted LaserDisc.

What this does not solve

Three limits, stated because the argument is weaker with them concealed and because each is a real research problem rather than an engineering oversight.

Re-embedding does not restore parity. Recomputing every vector under a new model gives you a working index, not the same index, and retrieval quality shifts in ways that require measurement rather than assumption. Keeping the source text converts a catastrophe into a project; it does not make the project free.

Distillation is lossy and carries the idiom of whichever model performed it. A system that summarises its own history into durable notes has improved its storage economics and introduced an editorial layer. Retaining the original observation contains the problem. It does not remove it, and there is no version of automated summarisation that is neutral about what matters.

And nobody has run any of this long enough to know. The published work on transferring agent memory between models reports encouraging continuity scores across three frontier models, on a pilot of fifty tasks.12 Fifty tasks is a promising indication and not a basis for a claim. The honest position is that the argument for keeping a model-neutral source of truth rests on reasoning about failure modes rather than on longitudinal evidence, because the practice is not old enough to have produced any.

Keep the source text

What the ada-002 notice exposed was not a flaw in OpenAI's deprecation policy, which was more generous than most. It exposed how many systems had stored the output of a process and discarded the input, and had done so without anyone deciding to.

That decision is usually made by omission. Nobody writes down that the transcripts can be deleted once the summary exists, or that the source text need not be kept once the vectors are built. It happens because storage costs money, the derived artefact is the one in daily use, and no failure occurs until the day the model changes.

The clerks of 1086 had no theory of digital preservation and no choice of format. The BBC in 1986 had both, and chose the medium that gave the better experience, which at the time it plainly did. What neither team could have been expected to do is what is now available to anyone building these systems: to keep the thing from which everything else can be rebuilt, in a form that does not require a particular machine, and to treat everything computed from it as replaceable however much it cost.

An organisation cannot know today which model it will be using in five years, and on current evidence should assume it will be several. The question worth putting to whatever is accumulating its context is not how much it remembers. It is what would remain if the model underneath it were withdrawn on twelve months' notice, and whether anyone has checked.

Footnotes

  1. OpenAI, API deprecations. text-embedding-ada-002 carries a deprecation date of 4 January 2025 with retirement no earlier than 14 June 2025, and the recommended replacements are text-embedding-3-small and text-embedding-3-large. The twelve-month notice period is longer than several comparable withdrawals elsewhere in the industry. ↩

  2. The BBC Domesday Project, November 1986. Around a million contributors, predominantly schoolchildren, surveyed their own localities; the material shipped on two LaserDiscs read by a BBC Master fitted with a specific interface and disc player. ↩

  3. The CAMiLEON project (Creative Archiving at Michigan and Leeds: Emulating the Old on the New) ran from 1999 to 2002, led by Margaret Hedstrom with Paul Wheatley managing, and recovered approximately 70 gigabytes per disc side by emulating the original system. A separate effort located the one-inch videotape masters. The National Archives holds the 1086 survey. ↩

  4. Qdrant, "Migrate to a New Embedding Model with Zero Downtime in Qdrant." The documentation states that switching models requires re-embedding all vectors in a collection, and assumes the source data is available in the payloads. ↩

  5. Anthropic's memory tool provides a file directory persisting across sessions; paired with server-side context editing, reported token savings reach approximately 84 per cent on long-running tasks. These figures are the vendor's own and should be read as such. ↩

  6. Anthropic has documented memory import and export, including a route for bringing context from other tools. No equivalent export API for accumulated user memories has been published by OpenAI as of August 2026. This reflects the public position at the time of writing and is the kind of gap that closes without announcement. ↩

  7. Rishi Jha, Collin Zhang and colleagues, "Harnessing the Universal Geometry of Embeddings," arXiv:2505.12540, presented at NeurIPS 2025. Subsequent work on scaling the approach with linear transformations appears as arXiv:2510.02348. The authors present the result as a security finding about information leakage from stored embeddings, and it is worth reading in those terms rather than as a migration tool. ↩

  8. The Modeled Information Format, version 1.0.0, June 2026, MIT licence. It describes itself as a content model for the Open Knowledge Format, which is a vendor-neutral markdown specification published by Google Cloud under Apache-2.0. ↩

  9. The W3C AI Agent Memory Interoperability Community Group, proposed 18 May 2026 and chartered 19 June 2026. Community groups are not standards-track bodies; their output is a report unless separately taken up. ↩

  10. draft-saihm-memory-protocol-01, Independent Submission stream, intended status Informational, expiring 28 November 2026. The document states explicitly that it is not endorsed by the IETF and has no standing in the IETF standards process. ↩

  11. Apache Ossie (incubating), formerly the Open Semantic Interchange initiative, which published a version 1.0 specification in January 2026 and entered the Apache incubator in June 2026 under Apache-2.0. ↩

  12. The pilot reports transfer continuity between 0.83 and 0.92 across three frontier models, against 0.28 to 0.45 without structured transfer, on a sample of fifty tasks. The authors themselves note that large-scale validation across hundreds of diverse tasks remains to be done. ↩