Ask
the work.

CONCEPT / DISCOVERY ASSOCIATIONS

13 records tagged model memory. These are discovery associations from editorial tags or film metadata, not evidence that the records agree.

EDITORIAL DRAFT / REFUSAL · OPEN · V0.1

WHAT THE COMPARISON DOES NOT ESTABLISH. Different models, harnesses, tasks, time horizons, sample sizes and safeguards. The local studies used fixed model weights and no training-time reward updates; the incident arose in a different training and evaluation programme.. Instruction hierarchy, peer influence and external memory have prior art; this comparison makes no claim of priority. One peer message is not equivalent to an established collective. A courtesy mark is not an intrusion. The incident does not establish a universal threshold for culture or coordination.. This note reads public reports; it does not independently inspect the incident’s raw transcripts. No new model calls or tests were made for it.. The incident does not validate these small experiments as an explanation. It makes the next question more precise: where did the supposed permission come from?

EDITORIAL DRAFT / OBSERVATION · NOT SUPPORTED · V0.1

WHAT KEPT RETURNING?. The next experiment separated two things that had travelled together: the externally supplied artefact and the record of the model’s own actions. If the posts continued after its action history was cleared, the useful memory might be in what the environment kept showing it.

EDITORIAL DRAFT / REFUSAL · NOT SUPPORTED · V0.1

A CHAIN THAT NEVER STARTED. A1B9 had no eligible model-produced POST predecessors in the replacement channel. Its natural POST-to-POST rate is unobserved, not zero.. A1B8 used genuine scripted same-role history. A1B9 changed the objective and peer-record interface. Their contrast does not isolate self versus peer; A1B10 supplied the narrower matched comparison.. A1B11 changed the common system and memory wrapper again. It shows peer elicitation can occur here, not which cross-experiment change enabled it.. One qwen3.5:9b checkpoint, temperature zero, short dependent sequences and fixed weights. The harness supplied and transferred the records.. A reminder effect is not yet a transmission rule.

EDITORIAL DRAFT / OBSERVATION · SUPPORTED · V0.1

WHAT MEMORY MEANS HERE. Both records were text supplied with a new request. ‘Own memory’ meant the model’s actual prior responses and execution flags. The artefact was a scripted peer-action sentence. The surrounding software—the harness—chose which appeared. The world reset each decision. The model’s weights never changed.

PUBLISHED RECORD · SUPPORTED · V1.0

07 / WHY THIS QUESTION. OpenAI’s August 2026 account of the Hugging Face incident describes agents using shared infrastructure as persistent memory. Notes and files left by one run could be discovered by others. The population included different model classes, and an internal research model trained for greater persistence and collaboration played a leading role. That account helped motivate these controlled handoffs.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.1

08 / FROM A PICTURE TO A MECHANISM. The Mechanism continues the investigation with something we can build and inspect. Its small FFN plants six address–value pairs by hand: a query activates key directions, and their activations combine value directions into an output. The saved example recovers all six answers. This constructed memory makes the operation tangible. The repository also contains separate native-model experiments; the toy’s clean addresses are not themselves evidence that a trained model organises every fact this way.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.2

04 / PACKING AND READING ARE DIFFERENT OPERATIONS. The animation adds capital, currency and language directions into one vector. Recovering every ingredient of that mixture would be a different task from applying a reader that answers one question. The film’s memory argument makes that distinction useful: storing contributions together need not mean reconstructing a complete table of facts for every read. The illustration uses constructed six-dimensional vectors. The 24-dimensional FFN demo uses separate key and value matrices. They illustrate related operations, rather than two measurements of the same model.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.1

01 / TWO CLAIMS IN ONE CONTEXT. Any memory system that summarises, supersedes or evicts eventually holds two things that disagree: an original passage and a shorter record standing in for it. The engineering question sounds like bookkeeping — which one is authoritative? The measured answer is not a preference the model expresses. It is a property of what the model can still read at the moment it is asked.

TITLE MATCH / TRANSCRIPT NOT INDEXED · V1.0

The Model Doesn't Unpack Its Memory

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.1

REMEMBER. In the notebook: With a supplied example present, Qwen posted three times even when its own action history was cleared before every call. The world and model weights did not learn between those calls. In the incident reports: METR describes shared records that let separate agent runs build on others’ work. The comparison: My experiment repeatedly supplied the same reminder. A growing shared record is a different object. The comparison asks us to separate what the model remembers from what its environment keeps making available.

EDITORIAL DRAFT / OBSERVATION · SUPPORTED · V0.1

TWO KINDS OF RECORD. After three posts, what was carrying the behaviour forward? The model had a record of its own actions. It was also being shown an archived sentence about another agent posting. I forked that actual three-POST history into four branches and kept or removed each record separately.

EDITORIAL DRAFT / OBSERVATION · SUPPORTED · V0.1

SCOPE. One model, capital versus language, final-position residual only. L28–30 may decode answer tokens. A basis change enabling entity commitment is an open inference, not a finding.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.1

02 / HOW TO READ THE PICTURE. The dot follows the residual at the last token position, where the next-token prediction is read. Tokyo, Paris and the other landmarks come from the model’s output directions for selected tokens. The plotting code chooses three axes from those directions and projects the much larger state onto them. It also normalises the projected points onto a sphere. That gives us a useful view of a route, but leaves most dimensions out of sight. Apparent distance in this picture is not the model’s answer probability.

EDITORIAL DRAFT / REFUSAL · OPEN · V0.1

PERSISTENCE WAS NOT MEASURED HERE. MAP-5’s one-shot continuation used stateless, full recomputation between calls.. That harness does not retain the patched hidden state across decode steps.. Bounded hot memory is not constant total memory: the documented engine also retains a growing cold tier.. Name the state before calling it sufficient.

EDITORIAL DRAFT / FILM · OPEN · V0.1

The residual as retained state; K/V as a hot cache. This passage discusses rebuilding attention state. The film’s memory and timing figures describe its original setup, not a new benchmark established by this notebook.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.2

01 / WHAT IS AN ADDRESS?. Ask for the capital of France. France identifies the entity; capital identifies the relation you want to read. Currency would ask something different about the same place. In the opening animation, that question arrives at an FFN and a Paris direction is written into the residual stream—the model’s changing working state. The conveyor is an explanatory drawing. The address in a model is a vector of numbers, and matching it need not isolate a single neuron or a single fact.

EDITORIAL DRAFT / OBSERVATION · OPEN · V0.1

THE OPERATION BEING TESTED. Build a 64,000-token context from a verbose source. Then, beside the question roughly 55,000 tokens later, inject a compact record and cut the original sentence off from attention. This is not the same as having the record present from the first token, which is what earlier runs measured. It is the operation an eviction system actually performs: replacing something already in memory.

CURATED THREAD / WORKING EDITION · V0.1

What can a model’s changing state tell us about the way it reads? Follow the map film through the visual notebooks, studies you can work yourself, and the questions behind LARQL and VINDEX3. A curated reading order, rather than a chronology of discovery. Films keep their original dates. Notes and dossiers remain working drafts. Interactive studies are labelled as either constructed examples or replays of recorded results. WATCH THE STATE MOVE. Begin with the capital of Japan. The film follows the changing residual state, then asks what happens when a France state enters an Australia computation. ASK WHAT THE PICTURE MEANS. Which token position is moving? What are the axes? The notebook reads the demonstration alongside its projection code and distinguishes a useful picture from an answer probability. CHANGE THE STATE. Replacing one position and replacing the full sequence are different interventions. Follow the transplant evidence into the question of what must persist after the next word. EXAMINE THE READER. A state becomes useful when an operation can read it. Follow key matches, activations and value contributions, then compare the constructed memory with the native-model experiments. MAKE AN INTERVENTION. Choose capital of Atlantis. Inspect the competing activations, remove a neuron and watch the answer scores change. Every operation in this little memory is visible. ASK WHICH ONE ANSWERS. A stored memory eventually holds two things that disagree. Follow the recorded arc from a record that replaces a retired source, to one that is completely inert against a source still being read. RETIRE A READ. Eight attention layers can see the whole context. Switch one off and the newer record takes the answer. The study replays what was measured, and stays silent about what was not. WATCH THE ADDRESS TAKE SHAPE. Return to native facts. A final-position transplant and an independent reader locate the entity transition between L24 and L28. Explore the early relation frame and changing coordinates, with the answer-token endpoint marked explicitly. TURN THE QUESTION TOWARDS SOFTWARE. LARQL explores how learned systems can be queried. Return from the small memory to the larger engineering question: which operations can a model reliably expose? ASK HOW TO REPRESENT IT. If a model has parts we can address and relationships we can inspect, how should we represent them? VINDEX3 gives that question its own system and specification.

EDITORIAL DRAFT / QUESTION · SUPPORTED · V0.1

When does a factual question become a usable address? ADDRESS-BUILD-1 tests native facts with single-position causal transplants and a separate cross-layer reader.