MnemoverseDocs

Appearance

Edward Izgorodin·Published Jul 11, 2026

Claude's Global Workspace: Why AI Memory Lives Outside the Model ​

Anthropic's interpretability team found a small, causally important working space inside Claude. Concepts enter it, influence other computations, and steer the answer.

Then the forward pass ends.

The space disappears. The model can voice what was in it, but cannot audit it. Nothing in the mechanism verifies whether its contents are true or carries them into the next session.

That distinction matters more for agent design than the metaphor of a machine "thinking."

TL;DR ​

  • Anthropic demonstrated a workspace-like structure in Claude that holds at most about 25 single-token concepts the lens resolves as active, and influences computations across the model.
  • Concept swaps and ablations show this workspace is causally involved in multi-step reasoning, not merely correlated with it.
  • The workspace exists within one feedforward pass. It does not provide durable, inspectable, or verified memory.
  • For AI agents, the three things memory must provide—persistence, ground truth, and inspectability—are architecturally absent inside the model. They don't emerge with scale; they have to be built outside.

What Anthropic actually found in Claude ​

A global workspace, as used in this paper, is a small set of concept representations that many parts of a language model can read from and write to during one feedforward pass.

In July 2026, Gurnee, Sofroniew, et al. published "Verbalizable Representations Form a Global Workspace in Language Models". Anthropic also released a plain-language overview titled "A global workspace in language models".

The researchers built a method called the Jacobian lens, or J-lens. For each model layer, it measures the average sensitivity of the final hidden state to intermediate activations, computed over a corpus of one thousand prompts sampled from a pretraining-like distribution. The resulting directions identify activation patterns with unusually strong causal influence on the output, and each J-lens vector corresponds to a single-token concept.

The researchers define J-space as "the set of points expressible as a sparse nonnegative combination of J-lens vectors." In plain terms, it is the small conceptual working set the lens exposes.

That working set is narrow. The paper reports a sparsity of "no more than 25, which we empirically observed to be the number of J-lens vectors that are meaningfully active at a given time." The overview describes this more loosely as only a few dozen concepts.

It also behaves like a communication hub. According to the overview, far more network components read from and write to these patterns than to ordinary representations—in some parts of the network, by a factor of about a hundred. The paper states the same phenomenon qualitatively: J-lens vectors compose more broadly than other representational vectors.

Connectivity alone would not prove the workspace matters. The causal interventions do.

In one experiment, the researchers asked the model to think of a sport, then subtracted the projection onto the Soccer lens vector and added an equal-magnitude projection onto the Rugby lens vector. The model reported "Rugby" as the sport it had thought of.

A second test redirected a two-step factual inference. The prompt was: "The number of legs on the animal that spins webs is". The model first had to infer spider, then recall the leg count, producing 8. When the researchers swapped the spider lens vector for ant mid-computation, the leading output changed to 6, an ant's leg count.

The ablation result was broader. Removing the strongest J-space directions dropped multi-step reasoning from "near-ceiling performance … to near zero," according to the technical paper. The effect was selective. MMLU, SQuAD, sentiment classification, and much ordinary text prediction stayed at or near baseline. Caesar-cipher decoding, analogy completion, summarization, TriviaQA, multi-hop reasoning, translation, and sonnet writing fell well below unablated Haiku 4.5.

So this is not a blanket "fact recall survives." Free-form factual recall on TriviaQA collapsed; the structured, extractive, and multiple-choice tasks were the ones that held.

The primary experiments used Claude Sonnet 4.5. The researchers corroborated key results on Haiku 4.5 and Opus 4.5, with some analyses conducted on Opus 4.6.

Claude's working memory lives for one breath ​

Working memory holds information needed for the current computation; long-term memory preserves information so it can be used after that computation ends. The paper demonstrates the first mechanism, not the second.

Claude's workspace operates "within a single feedforward pass rather than through recurrent loops," according to the paper; the overview says it evolves over a single pass through the network. When the pass ends, those activations are discarded and the weights remain frozen. Nothing in the workspace carries its active concepts across independent inference calls or sessions—unless something outside writes them into the next context.

Can the model see into this space itself? Partly—and the paper's own headline word marks the boundary: the representations are verbalizable. Ask Claude what it is thinking about, and it can report concepts active in J-space. But a report is not an audit. The self-report is transient and unverifiable; there is no channel through which the model can enumerate, check, or later retrieve its workspace as an inspectable record. That completeness exists only through the researchers' external J-lens. The thought is causally load-bearing—its owner can voice it, but cannot audit it.

The lens itself has limits. The authors say it "only approximately and incompletely captures the model's underlying workspace structure," and it detects only concepts represented by single vocabulary tokens, so phrases and even small integers can escape it. In roughly the first third of the model, readouts are noisy and largely uninterpretable—possibly a limit of the lens rather than an absence of workspace activity. The finding is substantial but bounded: researchers exposed and steered part of a transient working mechanism, not a durable internal store.

A mind without a memory — strictly as an analogy ​

The parallel that comes to mind is patient H.M., Henry Molaison—an analogy, not a claim that Claude and a human brain share the same architecture. In 1953, surgeon William Scoville performed a bilateral medial temporal-lobe resection that removed most of Molaison's hippocampus to treat severe epilepsy; Scoville and Milner (1957) documented the result. H.M. kept his intellect and in-the-moment working memory but could no longer form new long-term memories—every conversation vanished. The classic dissociation: thinking intact, remembering gone. Claude's workspace shows a limited computational parallel: it holds causally necessary intermediate concepts, yet nothing downstream keeps them.

The term global workspace comes from Bernard Baars's 1988 A Cognitive Theory of Consciousness, which described cognition as many specialist processes sharing information through a limited-capacity broadcast. Stanislas Dehaene's global neuronal workspace account ties conscious access to "ignition," a sharp, all-or-nothing amplification. Earlier AI work built such workspaces in by design—Goyal et al. (ICLR 2022) and VanRullen and Kanai (2021) each constructed a shared bottleneck for inter-module communication. What makes the Anthropic result different is that the workspace emerged unasked in a production model no one designed to have one. The researchers found it, read it, and steered it. They did not build it.

The paper is careful about what it does not confirm. It observes some competition for access to J-space, but says "it is unclear whether this mirrors the sharp, competitive 'ignition' that characterizes workspace entry in the brain." Partial competition, yes; brain-like ignition, not demonstrated. On the brain more broadly: "We do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain." And on phenomenal consciousness, the paper is direct: "we take no position on this issue." Anthropic's overview adds that the experiments "don't show Claude can have experiences, or feel things in the way humans do."

Why AI agent memory has to live outside the model ​

The paper demonstrates an internal working mechanism. The external-memory conclusion is an engineering argument that follows from its boundaries.

An AI agent memory system needs three properties this workspace does not provide:

  1. Persistence. Workspace activations disappear with the forward pass, and frozen weights record no new event. Nothing inside carries state across calls or sessions.
  2. Ground truth. The workspace runs on whatever concepts enter it, accurate or invented. There is no reliable internal mechanism that checks those concepts against an authoritative record; the model can reason fluently from a false premise.
  3. Inspectability. The model can voice what is active in its workspace, but it cannot audit it—there is no way for it to enumerate or verify the contents, and only an external interpretability tool can expose them completely.

These three are absent by construction, not merely undiscovered—and here the argument rests on the architecture, not on the lens. The J-lens is, by its authors' own account, an approximate and incomplete view of the workspace, and it may yet miss concepts the model uses. But persistence, grounding, and inspectability do not depend on what the lens can see. Discarded activations and frozen weights store nothing across passes; a network that can voice its active concepts but not enumerate or verify them cannot audit itself. Scale can produce richer computation inside a pass—the workspace itself apparently emerged without explicit design—but it does not turn discarded activations into persistent records or create an external source of truth. That is why external memory exists as an architectural category: it retains knowledge after inference ends, connects claims to records, and gives systems something inspectable to retrieve later.

The small working set also offers a practical design hint. With at most ~25 lens-resolved concepts meaningfully active at once, the scarce resource is the model's attention of the moment, so external memory systems gain more from careful selection of what to surface than from raw volume.

The model handles the moment. The surrounding system must preserve what the next moment needs.

Common questions ​

Did Anthropic find consciousness in Claude? ​

No. The paper explicitly states, "we take no position on this issue." The experiments study the functional role of a limited-capacity broadcast system, not phenomenal experience. The overview adds that the experiments "don't show Claude can have experiences, or feel things in the way humans do."

What is a global workspace in an LLM? ​

In this research, a global workspace is a set of highly influential activation patterns—found with a new tool called the Jacobian lens—that behave like a central hub. Far more network components read from and write to these patterns than ordinary ones, and they carry causally load-bearing concepts that the model uses for multi-step reasoning.

Does Claude remember between conversations? ​

No. The workspace lives inside a single forward pass. Activations are discarded when the pass ends, weights are frozen, and no mechanism inside the model carries state across calls or sessions. The model can voice what is active in its workspace, but it cannot audit or later retrieve it—complete visibility requires external researcher tools.

Why can't the model just remember things itself? ​

On our reading of the paper's boundaries, three requirements for memory—persistence, ground truth, and inspectability—are architecturally absent inside the network. Persistence would need state that survives the forward pass; ground truth would need a reliable check against an authoritative record; inspectability would need a way for the model to audit its own contents, not just voice them. None of these emerge from scaling; they must be built outside.

What does this mean for AI agents? ​

Agents need memory that persists across calls, stays grounded in facts, and can be inspected. The workspace finding shows that even a frontier model's internal "thought" is transient—voiceable in the moment, never auditable afterward. External memory layers—built for persistence, grounding, and auditability—are the practical engineering answer, not an optional add-on.

Is the workspace the same as human working memory? ​

The researchers explicitly state, "We do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain." The analogy is useful—a limited, causal buffer that holds a few concepts at once and then vanishes—but it is an analogy, not a claim of brain equivalence.

Edward Izgorodin · Mnemoverse · last updated 2026-07-11