Skip to content
Menu
Programme

Representation and memory

Minds and models carry internal state and stored knowledge that decide their behaviour but cannot be read directly. This programme asks what can be read from outside, which readings reflect what a system uses, and how a memory can be revised and checked as the world changes.

What does a model hold, which of it does it use and report, and how does what it holds stay correct as the facts it rests on change?

Generated cyanotype print of a nautilus shell cut cleanly in half, showing its spiral of sealed chambers.Generated photograph of a white sectional model of a three-storey house showing its rooms, with vermilion beads on small objects in three rooms.Generated render of an iceberg in cross-section, its small cap above the waterline and its large hidden body below.
Plate IA shell cut open: its growth is kept inside as sealed chambers.An interior opened by a section. What is kept inside can be marked, but only from outside.Most of it lies below the line, known only from outside. Interiors are measured, not seen.Generated image in a cyanotype treatment. Far Akin art direction.Generated image. Far Akin art direction, Datum look.Generated image. Far Akin art direction, Swath look.

Why it matters

A model's behaviour is set by what it holds: representations formed during computation, associations stored in its weights and, in agents, records kept outside the model across sessions. Oversight and correction both depend on reaching that interior from outside. Recent work reports that language models hold a small set of representations they are poised to verbalise, which show functional properties of a global workspace and structural signatures that workspace theory associates with conscious access in people[1][2]. If readings of that kind hold up, a model's stated reasoning could be checked against the structure that drives its behaviour. Such readings would not show that a model is conscious. The inquiry tests whether the readings repeat, not whether a model has experience.

Memory raises the same need over time. The information stored in a language model falls out of date quickly[3]. Agents that keep long-term records must handle knowledge updates and temporal reasoning, and chat assistants and long-context models have shown marked drops in accuracy across sustained interactions[4]. Systems such as MemGPT manage tiers of memory to extend what a model can draw on[5]. Whether such stores stay correct as the world moves is an open question.

What is hard

Readings of an interior are easy to over-interpret. A probe can score well because the probe has learned the task, not because the representation encodes it[6]. A subspace intervention can change a model's output as if a feature had changed while acting through a separate pathway[7]. A measure of an interior is worth what its reliability and validity are worth, and both have to be shown.

Editing stored knowledge is harder than it first appears. Factual associations in some models correspond to localised computations that can be edited directly[8]. But an edit to one fact implies changes to others. Editing methods recall the edited fact yet fail on questions whose answers should change as a consequence[3], and fail to make consistent changes across related facts[9]. In both studies, methods that kept the edit out of the weights (an external memory in one, the prompt in the other) did better.

Memory over time adds a further problem. People change their minds, and a store that treats every statement as a fixed fact cannot represent revision. A benchmark for belief dynamics in multi-session interactions reports a trade-off: models that personalise aggressively resist drift poorly, and factually grounded models miss legitimate updates[10]. Complementary learning systems theory proposes that the brain pairs a fast store for new items with a slow system that learns structure gradually, so that new learning does not disrupt what is already organised[11]. Which structure an artificial memory needs, and which layers earn their cost, has not been settled.

Current directions

  • Instruments before theory. Testing whether workspace measures repeat across refits and seeds, and whether different measures agree, before any model is trained to have a workspace.
  • Breadth of change. Measuring how editing methods behave as the number of facts that depend on one changed fact grows.
  • Foreseeing failure. Asking whether a model can forecast which later interference will break a stored memory, building on evidence that language models can estimate whether they know an answer[12].
  • Structure that earns its cost. Comparing memory designs one layer at a time, from full context and raw retrieval to typed, time-aware and revisable records.

How the lab approaches it

Four inquiries share the area. Can the instruments for a model's workspace be trusted? would apply three workspace tests to published open-weight models and ask whether their readings repeat and whether they rank models alike. When one fact changes would measure how far dependent facts update, and unrelated facts shift, as the number of dependents of one changed fact rises. Can a memory foresee what will break it? asks whether a trained predictor forecasts interference with a stored memory better than a model given the same information without training on outcomes. What a memory should keep apart asks whether agent memory should keep claims, facts, beliefs, contradictions and superseded items as separate states, with their source and time recorded. It would set each richer memory against the thinner one beneath it at matched cost, and keep a layer only if it beats that memory on a hard case by a margin to be fixed in advance and has a mechanism that explains the gain.

Each design states what a pass would not show. Stable readings are not evidence that a model resembles a mind, and a curve measured on one set of changed facts is not a limit on editing in general. The lab treats any resemblance between a model's interior and human cognition as a question to test, not a premise.

The longer aim is a set of instruments and memory designs that let what a model holds be measured and corrected for a stated purpose.

Open problems in this programme

References

  1. Dehaene, S., Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework. Cognition 79(1–2), 1–37.
  2. Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., et al. (2026). Verbalizable Representations Form a Global Workspace in Language Models. arXiv preprint 2607.15495.
  3. Zhong, Z., Wu, Z., Manning, C. D., Potts, C., Chen, D. (2023). MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions. Proceedings of EMNLP 2023, 15686–15702.
  4. Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., Yu, D. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. International Conference on Learning Representations (ICLR 2025).
  5. Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., et al. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv preprint 2310.08560.
  6. Hewitt, J., Liang, P. (2019). Designing and Interpreting Probes with Control Tasks. Proceedings of EMNLP-IJCNLP 2019, 2733–2743.
  7. Makelov, A., Lange, G., Nanda, N. (2023). Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching. arXiv preprint 2311.17030.
  8. Meng, K., Bau, D., Andonian, A., Belinkov, Y. (2022). Locating and Editing Factual Associations in GPT. Advances in Neural Information Processing Systems 35 (NeurIPS 2022).
  9. Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M. (2024). Evaluating the Ripple Effects of Knowledge Editing in Language Models. Transactions of the Association for Computational Linguistics 12, 283–298.
  10. Myakala, P. K., Agrawal, M., Manche, R. (2026). BeliefShift: Benchmarking Temporal Belief Consistency and Opinion Drift in LLM Agents. arXiv preprint 2603.23848.
  11. McClelland, J. L., McNaughton, B. L., O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological Review 102(3), 419–457.
  12. Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., et al. (2022). Language Models (Mostly) Know What They Know. arXiv preprint 2207.05221.

Inquiries in this programme

  • Generated cyanotype print of three weighted lines lowered into one dark pool, their rings of ripples overlapping but not quite matching.Generated photograph of three different white probes lowered into one plaster basin, each tip marked by a vermilion bead; the three beads land near each other but not together.Generated night-survey render of three side-by-side strips of mapped sea floor that show the same ridge, nearly aligned.
    Designed

    Can the instruments for a model's workspace be trusted?

    On open-weight language models that already exist, do three tests of an internal “workspace” give repeatable readings (test-retest r of at least 0.8), and do they rank the same models alike or differently?

    Latent space and internal cognition · Cognitive science and psychology · Machine-learning foundations

  • Generated cyanotype print of a small stone cairn on a flat beach with a wave line advancing towards it.Generated photograph of a cairn of white plaster blocks on a plaster shore with terraced plaster sea rising toward it; one low block carries a vermilion bead.Generated render of a sandbar on the sea floor with two currents of points approaching it from either side.
    Designed

    Can a memory foresee what will break it?

    Given a stored memory and descriptions of two possible later disruptions, can a trained model predict which one will leave the memory unable to support its rule, more accurately than a model given the same information and no training on outcomes beyond a single fitted temperature?

    Memory, knowledge and continual learning · Cognitive science and psychology

  • Generated cyanotype print of a fishing net lifted at a single knot, the whole mesh hanging from that point.Generated photograph of a white knotted net on a plaster plinth, one central knot lifted on a vermilion thread, the nearby knots marked with vermilion beads.Generated render of a sea-floor surface where one raised point sends fading rings of change outward.
    Designed

    When one fact changes

    For a weight-based editing method at a stated update budget, does correct updating of dependent facts fall, and do wrong changes to unrelated facts rise, as the number of dependents of one changed fact grows?

    Memory, knowledge and continual learning · Machine-learning foundations

  • Generated cyanotype print of pressed leaves of different ages layered together, the oldest pale and skeletal, the newest dark and whole.Generated photograph of stacked white paper strata, the lower layers pushed aside by newer ones, pierced by a vertical vermilion thread that bends at the displaced layer.Generated render of a cross-section of stacked sea-floor sediment layers, with one folded and broken section.
    Designed

    What a memory should keep apart

    At matched model, token and latency budgets, does an agent memory that stores claims, facts, beliefs, contradictions, supersession and uncertain freshness as separate states produce more correct later answers and actions on published benchmarks of changing information than full context, raw-event retrieval or typed records alone?

    Memory, knowledge and continual learning · Measurement and research methodology