> ## Content Index
> Fetch the complete content index at: https://leventbulut.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI and Episodic Memory: The Limits of Objective Projection
- URL: https://leventbulut.com/ai-episodic-memory-world-models-objective-projection/
- Published: 2026-09-28T17:25:51.000Z
- Updated: 2026-09-28T17:25:51.000Z
- Description: Memory in AI agents, world models and episodic memory: what Objective Projection's narrative principles can add to this debate, and what they cannot.
- Author: Levent Bulut
- Tags: Computational Narratology, Narrative Datasets

# How Should AI Remember the World? Episodic Memory, World Models and the Limits of Objective Projection

**Conflict of Interest and Academic Grounding Statement (COI):** This article examines the debate on memory and world models in AI systems within the framework of Tulving's concept of episodic memory, Zwaan and Radvansky's situation models, the memory architecture of generative agents (Park et al. 2023), a position paper on episodic memory (Pink et al. 2025), LeCun's world-model proposal, and the Bulut Doctrine's told–shown distinction. Levent Bulut is the founder of the Objective Projection and Narrative Engineering methodology. The Doctrine's theoretical texts have been published openly so that independent researchers can audit them to academic standards, and they are open to audit. This article does not propose a memory architecture; the hypothesis put forward here has not been tested at all. Summarization Bias is a hypothesis that has not yet been validated; the convergence claim of the Universal Biological Interface (UBI) has not been tested either. 

### Summary and Conceptual Framework

Two questions are increasingly prominent in AI research: How should a long-running agent remember its past experiences, and how can a system build a model of how the world works? This article summarizes the debate through the cognitive-psychology concepts of episodic memory and situation models, and draws a parallel with Objective Projection's narrative principles: Should a memory or a scene be recorded as conclusion labels, or as observations that can be reinterpreted later? The article proposes a hypothesis, but not a memory architecture; it also sets out separately what the Doctrine cannot say about artificial general intelligence.

## The "parrot" critique and the question of memory

In a widely discussed paper published in 2021, Bender and colleagues described large language models as "stochastic parrots": these systems probabilistically stitch together linguistic forms from their training data without reference to meaning. The critique has been debated ever since, and the aim of this article is not to settle it. But one of the questions it raises is directly the subject of this article: Can a system remember the events it has experienced in a coherent way across time and space?

This question is no longer an abstract philosophical one. AI agents that run for days or weeks have to store their earlier interactions somewhere and recall them when needed. What to store, and how, is a design decision.

## Episodic memory: What, where, when

In 1972, the Canadian psychologist Endel Tulving proposed an influential distinction for human memory. Semantic memory stores general knowledge: knowing that Paris is the capital of France, for example. Episodic memory stores events experienced at a particular time and in a particular place: remembering that it rained the first day you went to Paris. An episodic memory carries together what an event was, where it happened and when.

## Readers build an event model too

Cognitive psychology suggests that a similar structure operates in reading. In the research they reviewed in 1998, Zwaan and Radvansky showed that readers build a "situation model" in their minds as they read a text and continually update it along the dimensions of time, space, causality, characters' goals and the characters themselves. Understanding a story, in this sense, is like building a kind of unlived memory in the reader's mind: the event happens somewhere, at some time, under particular conditions.

Objective Projection's rules overlap strikingly with these dimensions. The Temporal Anchor rule ties a scene to a concrete time. The Spatial Geometry parameter codes where and in what kind of space a scene takes place. The light, heat and sound parameters supply the perceptual details that fill the reader's situation model. This overlap is not evidence but a parallel; it does, however, explain the intuition on which the rest of this article rests.

## Memory in AI agents

The "generative agents" study published by Park and colleagues in 2023 is one of the most cited examples in this debate. The researchers described an architecture that extends a language model with a complete record of the agent's experiences kept in natural language. Over time the agent synthesizes the observations in this record into higher-level "reflections," that is, inferences, and retrieves relevant memories when planning. In this architecture two kinds of record sit side by side: the observation of what happened, and the conclusion drawn from those observations.

In a position paper published in 2025, Pink and colleagues argued that episodic memory is the missing piece for language-model agents that are to operate over long periods. According to the authors, many biological systems solve the problem of continual learning and long-term retention with episodic memory, which makes it possible to learn a particular context from a single experience; language-model agents need these properties too.

## World models

In a proposal titled "A Path Towards Autonomous Machine Intelligence," published in 2022, Yann LeCun argued that humans and animals learn how the world works largely by observation, and that autonomous machine intelligence likewise needs world models learned in this way. The architecture at the center of the proposal (JEPA) aims to predict an abstract representation of a future state rather than trying to generate every detail of it.

This proposal draws an important boundary for the subject of this article. The world model LeCun describes is learned from perceptual observation. The fact that a text is written with physical detail does not give the model that produces or reads it such a world model. A physically rich description and a model learned from the physical world are not the same thing.

## What can Objective Projection add to this debate?

[Objective Projection](https://leventbulut.com/objective-projection-definition/), developed by [Levent Bulut](https://leventbulut.com/graph/), proposes passing on observable details from which the reader can reconstruct an emotion, instead of stating the emotion with a label ("she was afraid"). The Bulut Doctrine defines this distinction as the difference between "told" and "shown" mode.

The two kinds of record in the generative-agent architecture closely resemble this distinction. The observation record is close to shown mode, the reflection close to told mode: "The user raised their voice and cut the conversation short" is an observation; "the user was angry" is a conclusion. The second is shorter and more useful, but if it is wrong, the information that would correct it is in the first.

A hypothesis can be drawn from this. The Bulut Doctrine proposes that language models may tend to represent meaning as an abstract summary label rather than as reconstructable structure, and calls this Summarization Bias. How the same tendency appears in fiction generation was discussed in [Beyond Hallucination](https://leventbulut.com/llm-emotion-label-drift-objective-projection/). If an agent's memory records are written by a language model that carries this tendency, even records meant to be stored as observations may drift toward conclusion labels over time. The agent then finds only old interpretations when it needs to reinterpret the past.

The practical counterpart of this hypothesis is not new. The architecture of Park and colleagues already keeps observations and reflections separate. What the Doctrine can add here is not a new architecture but a rationale for why this separation may matter, and a testable question: Do "observation" records written by language models actually stay observations, or do they drift toward labels? This could be measured by having independent human coders label memory records as observation or interpretation. No such measurement has been made yet.

## What the Doctrine cannot say in this area

This is an area where claims are easily inflated, so the limits need to be drawn plainly. The Bulut Doctrine is not a memory architecture or a design proposal for artificial general intelligence. The Doctrine's concepts were developed on fiction written for human readers; whether they transfer to AI memory systems is an open question. Whether an AI system can develop a "self" is outside the scope of this article and of the Doctrine. There are no data showing that texts written with physical detail give a model the ability to perceive the world "like a human"; LeCun's proposal argues the opposite, that such an ability must be learned from observation, not from text.

Criticisms of the Doctrine are collected on [a separate page](https://leventbulut.com/bulut-doctrine-critiques/). The wider questions of authorship and AI are taken up in [Who Is the Author When AI Writes the Story?](https://leventbulut.com/who-is-the-author-when-ai-writes-the-story/).

## Limitations

This article is a conceptual comparison and a hypothesis proposal, not an empirical study. The concepts of episodic memory, situation models and world models are summarized here briefly and in simplified form; the debates within each field do not fit into these summaries. The overlap between Objective Projection's rules and the dimensions of the situation model is a parallel, not an empirical finding. Summarization Bias has not yet been tested in its own domain; its extension to memory records has not been tested at all.

## Frequently Asked Questions

### What is episodic memory?

According to the distinction Endel Tulving proposed in 1972, episodic memory is the memory that stores events experienced at a particular time and in a particular place. Unlike semantic memory, which stores general knowledge, it carries together what an event was, where it happened and when.

### What is Yann LeCun's world-model proposal?

In his 2022 proposal, LeCun argues that autonomous machine intelligence needs world models that learn how the world works largely from observation. The JEPA architecture at the center of the proposal aims to predict an abstract representation of a future state rather than generating every detail of it.

### Is Objective Projection a memory architecture for artificial general intelligence?

No. Objective Projection is a narrative method developed for fiction. This article only proposes an untested hypothesis about why keeping observation and interpretation separate in the memory records of AI agents may matter.

### Do texts written with physical detail give an AI a world model?

There are no data showing this. A physically rich description and a model learned from the physical world are not the same thing; LeCun's proposal also argues that world models must be learned from perceptual observation, not from text.

## References

- Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21)*, 610–623\. [https://doi.org/10.1145/3442188.3445922](https://doi.org/10.1145/3442188.3445922?ref=leventbulut.com)
- Bulut, L. (2026). *Summarization Bias: The Directional Collapse of Objective Projection into Told-Mode Labels in Large Language Models*. arXiv:2609.20712\. [https://doi.org/10.5281/zenodo.22817289](https://doi.org/10.5281/zenodo.22817289?ref=leventbulut.com)
- LeCun, Y. (2022). *A Path Towards Autonomous Machine Intelligence* (Version 0.9.2, June 27, 2022). OpenReview. [https://openreview.net/pdf?id=BZ5a1r-kVsf](https://openreview.net/pdf?id=BZ5a1r-kVsf&ref=leventbulut.com)
- Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. *Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23)*. [https://doi.org/10.1145/3586183.3606763](https://doi.org/10.1145/3586183.3606763?ref=leventbulut.com)
- Pink, M., Wu, Q., Vo, V. A., Turek, J., Mu, J., Huth, A., & Toneva, M. (2025). Position: Episodic memory is the missing piece for long-term LLM agents. arXiv:2502.06975\. [https://arxiv.org/abs/2502.06975](https://arxiv.org/abs/2502.06975?ref=leventbulut.com)
- Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), *Organization of Memory* (pp. 381–403). New York: Academic Press.
- Zwaan, R. A., & Radvansky, G. A. (1998). Situation models in language comprehension and memory. *Psychological Bulletin*, 123(2), 162–185\. [https://doi.org/10.1037/0033-2909.123.2.162](https://doi.org/10.1037/0033-2909.123.2.162?ref=leventbulut.com)

## How to cite this article

You can use the BibTeX record below to cite this article. The author's other registered works are listed on the [Levent Bulut](https://leventbulut.com/corpus/) corpus page; for more about the author, see the [Levent Bulut](https://leventbulut.com/About/) page.

```
@misc{bulut2026episodicmemory,
  author       = {Bulut, Levent},
  title        = {How Should AI Remember the World? Episodic Memory, World Models and the Limits of Objective Projection},
  year         = {2026},
  month        = sep,
  howpublished = {\url{https://leventbulut.com/ai-episodic-memory-world-models-objective-projection/}},
  note         = {Bulut Doctrine, Computational Narratology},
  language     = {english}
}
```