Prediction Is Memory Running Forward

By Erhan Bilal, PhD - CSO, Enkira AIMay 11, 2026

world modelsAI agentspredictionmulti-agent systemstheory of mind

Future agent

Part II: The Same Problem, Facing Forward

Why fluent agents still can't anticipate

Picture a coding agent making a small change to a shared configuration file. It reasons about the change, justifies it in a clean four-paragraph note, gestures at potential risks without sounding alarmist, and only afterward discovers that the other agent in the same session had spent the last hour writing three new files on the assumption that the original value would never move.

The note is, in its own narrow terms, a triumph: impeccable grammar, plausible reasoning, the right small concession to the people who might object. If you were grading prose, you would probably ask for the agent's contact information. The trouble is what the note does not contain. The system had every relevant fact about the actual state of the world. What it lacked was a representation of what the other agent believed about the world, and any machinery for treating that belief as something that could cause action.

The narrow phrase for what just happened is: fluency is not foresight.

Yann LeCun has argued, in various forms, that an intelligent agent needs a world model: some internal machinery for predicting how the world will change, including how it will change as a consequence of the agent’s own actions. That demand is right, but it is only the beginning. In a shared environment, intelligence is not just predicting what the world will do. It is predicting what other agents will do from the world as it appears to them, and how each action will change the memories they act from next.

A rat doing offline planning while looking like it has clocked out for the day

It is the mid-2000s. A rat finishes running a track and settles in the home box, doing rat things between trials. To external appearances, nothing of cognitive interest is happening. Inside its hippocampus, however, a sequence of place cells fires in compressed time, replaying the path the rat just ran, but backward, at roughly twenty times real speed. Foster and Wilson reported the result in 2006, with the practiced understatement of people who know they have just stepped on a tripwire in cognitive science.

Seven years later, Brad Pfeiffer and David Foster watched the same machinery run forward. At a moment of decision, with the rat at a junction, hippocampal place cells produced sequences sweeping out toward the goal location before the animal moved. Same hardware, different direction. Filing yesterday and rehearsing tomorrow, in the rat brain, are not separate organs in conversation with each other; they are postures of the same act, run by one sequence machine in different modes.

Daniel Schacter and Donna Rose Addis spent years arguing for the strong version of this. Human memory is reconstructive, lossy, and prone to confabulation, they said, not because evolution failed to make better recorders but because being a good recorder would have been worse for the thing memory was actually for. A perfectly veridical episodic system would be bad at every job memory actually has to do: generalization, recombination, prediction. The "imperfections" are the design. Or, less politely: human memory is bad at being a transcript because it is busy being a forecast.

The brain doesn't even replay everything. Howard Eichenbaum spent his career arguing that the hippocampus is more accurately understood as a map of relations, time, and problem structure than of physical space, and that replay over that map is selective. By the time Marcelo Mattar and Nathaniel Daw turned to the question in 2018, the selectivity itself was empirically settled; what they added was a normative account of why the selection looks the way it does. The brain prioritizes replaying the experiences whose review is most likely to improve future choice. Replay is less a recording played back than a planner deciding which fragments of the past are worth spending cognition on.

Useful memory, by this account, is compressed regularity weighted by significance. An agent doesn't merely need to remember that a migration failed last quarter; it needs the scar tissue, the embarrassment, the half-remembered sense that the failure was hard to roll back and traceable to a hidden assumption that, almost certainly, is still in the codebase. Memory of that kind constrains not just what future seems likely but which one deserves the cognitive spend.

The hard case: other minds

Single-agent prediction is hard, and multi-agent prediction is not just more of it. To anticipate what another agent will do next, you need a model of how that agent updates its expectations, which is to say a model of its memory. The cleanest failure mode of current agent stacks is false belief. Agent A predicts what Agent B will do based on the world as it actually is. Agent B continues to act from a stale picture. Their trajectories diverge, and Agent A has no way to represent the divergence, because nothing in its memory takes another agent's belief as a thing in itself. Retrieving the real state of the world cannot stand in for retrieving another agent's stale model of the world; those are different categories of fact. Retrieval-augmented generation can give you the first; it was never designed to give you the second.

The honest version of the problem is harder still, because beliefs are not static. Other agents are not passive variables in a state vector; they update on what they observe. If I act based on my model of you, you update your model of me, and the moving target I was trying to predict has just moved again. Sometimes my action corrects an earlier mistake of yours; sometimes it consolidates one; sometimes you read my correction itself as evidence that I am the kind of agent who corrects loudly, and you start filtering my next twenty messages through that. A message in a shared environment functions as an intervention on another agent's memory, not as an output token. A configuration edit is a belief update in every collaborator who notices it, and a stale-belief trap for every collaborator who doesn't.

To put it more starkly: in any shared environment, what other agents believe is not decoration on top of the state but a piece of the state itself, and the piece that determines what happens next.

What memory was for

The architecture that would let a system handle any of this does not yet exist. What it would need, more demanding than indexed storage of one's own past, is a registry of memories distinguished by whose world they describe and how that world tends to change under action: what I have come to know; what you appear to believe; what the team has come to treat as law; what the codebase actually does when it gets touched here. And because shared worlds change as soon as anyone acts inside them, those memories would have to track not just states but the way each of those states gets updated.

That is open research. It is not in any current agent stack, including the one this essay is being written for. The honest claim is much smaller, and it follows directly from Part I: most of what passes for "agent memory" today is a transcript and a search engine, neither of which is doing the work the brain has been doing for as long as there have been mammals to do it in.

Prediction is memory running forward. An agent that cannot remember cannot anticipate.

Part II: The Same Problem, Facing Forward

Why fluent agents still can't anticipate

A rat doing offline planning while looking like it has clocked out for the day

The hard case: other minds

What memory was for

Further reading