The Memory Problem: Why AI Video Models Forget Your World (and How to Fix It)

Make a shot. It is beautiful. Your protagonist looks exactly right.

Close the tab. Come back Thursday. Make the next shot in the same scene.

Her jaw is subtly wrong. The jacket has one fewer buckle. The light in her apartment now comes from the other side of the room, and the city outside the window has developed a colour palette it did not have on Tuesday.

Nothing broke. No error message. The model did what it was built to do, which is generate an excellent image of a woman matching your description. It just has no idea it has met her before.

This is the memory problem, and it is the single largest gap between AI footage and AI filmmaking.

Let us be precise about the claim

It is fashionable to assert that "everyone agrees memory is the biggest unsolved problem in AI video." We are not going to make that claim, because we could not find a defensible source for it. There is no survey, no named industry authority, no trade-press consensus you can point to.

What there is, and what is more useful, is a cluster of 2026 academic work that treats the problem as formally open, and builds benchmarks to measure it:

• EntityBench (arXiv, May 2026) evaluates entity consistency across long-range, multi-shot video generation. The fact that this benchmark needed to exist tells you the field does not consider the problem solved.

• MBench (arXiv, May 2026) measures memory capability in video world models specifically.

• Memento (arXiv, June 2026) proposes reconstruction-based approaches to consistent long video generation.

Three benchmarks and methods papers, all inside two months, all pointed at the same hole. That is a stronger signal than any vendor blog post, and it is honest about what it is: an active research frontier, not a solved problem with a product attached.

What the models actually give you

Every major model now ships consistency features. It is worth being exact about what they cover.

Seedance 2.5 (July 31, 2026; on Runway's API August 7) accepts up to thirty image references, ten video references, and ten audio references, generating four to thirty seconds in a single pass. That is the most generous reference budget available.

Kling 3.0 (February 2026) added subject-reference upload, multi-shot continuity within a sequence, and native audio.

Veo 3.1 offers reference-to-video and first-and-last keyframe control.

Runway Gen-4.5 (December 2025) inherited Gen-4's reference-image system and layered native audio and multi-shot sequencing on top.

Now the crucial distinction. Every one of those is within-generation consistency. You supply references, the model honours them for this output, and the arrangement ends when the generation ends.

There is no mechanism in any of them by which the model, unprompted, next Tuesday, knows that your protagonist has a scar on her left hand because of something that happened in scene six.

World models are a different thing (and not this thing)

Two developments get conflated with memory and should not be.

Runway GWM-1 (December 11, 2025) introduced a family of general world models: autoregressive, real-time, action-conditioned, built on Gen-4.5. Google DeepMind's Genie 3 went to research preview in August 2025 and reached AI Ultra subscribers via Project Genie in January 2026.

These are extraordinary systems. They simulate persistent, explorable environments in real time, holding physics and space coherent as you move through them.

What they hold is spatial and physical persistence. What a film needs is narrative persistence: canon. Who this person is. What she did in act one. What the rules of this world are and which of them she is about to break. A world model can keep a room from rearranging itself while you walk across it. It cannot tell you the room belongs to a character who has not lived there since her sister died.

Those are different problems. Solving the first does not solve the second, and the conflation costs filmmakers real time when they evaluate tools.

Why the drift happens

Three mechanisms, roughly in order of how much damage they do.

References are suggestions, not constraints. A reference image biases a generation. It does not bind it. Small deviations are within tolerance for the model and invisible per-shot, which is exactly why they accumulate. Nobody notices shot four. Everybody notices shot forty next to shot one.

Prompts drift because humans drift. The prompt you wrote in June and the prompt you write in August describe the same character differently, because you are a person and you have thought about her since. Every rewording is a small instruction to change her.

Nothing arbitrates. In a traditional production, the script supervisor is a job. Somebody's entire responsibility is noticing that the coffee cup was in the other hand. In most AI workflows, that role is unfilled, and the model is structurally incapable of filling it.

The working method

This is what actually closes the gap today, in rough order of leverage.

1. Build reference sheets, not reference images. One canvas, multiple views: front, three-quarter, profile, back, plus two lighting conditions and one expression range. A single flattering portrait is the most common mistake in AI character work, because it gives the model one angle to interpolate from and near-infinite freedom everywhere else. A 3x3 multi-view sheet constrains the whole space.

2. Storyboard in grids, not in singles. Nine panels generated on one canvas share a style pass by construction. Nine panels generated separately do not, and no amount of prompt discipline will fully fix it. This is the cheapest continuity win available and almost nobody does it.

3. Write the world bible before you write the scenes. Characters with history, not just descriptions. Locations with light sources named. A palette with hex values. Rules the world obeys. This document outranks any individual prompt, and it is the only artefact that survives a six-week gap in production.

4. Lock the palette explicitly and numerically. "Warm and cinematic" drifts. "Dominant burnt orange #ff5b10, deep black shadow, single warm practical raking from frame left" does not, or drifts far less.

5. Version your canon. When the world changes deliberately, record that it changed and when. Undocumented change and drift are indistinguishable six weeks later, and the difference matters enormously when you are trying to work out which shot is wrong.

6. Keep continuity outside the tool. Sora's shutdown is the cleanest available argument. OpenAI announced it in March 2026; the app went dark in April; the API ends in September. Disney had signed a roughly $1B deal to put its characters there and terminated it. Anything you have that only exists inside a vendor's product is on that vendor's timeline, not yours.

What structural memory looks like

The method above works. It is also, honestly, a lot of manual bookkeeping, and it degrades exactly when a production gets long enough to need it most.

The structural version is a living world: a persistent story universe where characters, locations, palette, and visual rules are defined once and enforced automatically on everything generated inside it. The canon is not a folder you maintain alongside the work. It is the thing the work is generated from.

On Magic Lantern, that layer is the AI Showrunner. It does what a script supervisor does, at the scale AI production actually runs at: holds the world's canon across every scene, every session, and every episode, so scene forty agrees with scene one without you re-supplying anything.

That is the whole design premise. Not better frames. Frames are close to solved. The unsolved part is everything that has to stay true between them.

What filmmakers are asking

Why do AI video models forget my characters between sessions?

Because reference features operate within a single generation. Models like Seedance 2.5, Kling 3.0, Veo 3.1, and Gen-4.5 will honour references you supply for that output, but they retain nothing about your project afterwards. Each new session starts with no knowledge of your world.

Is character consistency solved in 2026?

Not across a production. Within-generation consistency has improved substantially, with reference budgets reaching thirty images on Seedance 2.5. Cross-session consistency remains an open research problem, with 2026 benchmarks like EntityBench and MBench built specifically to measure how far from solved it is.

Are world models the answer to continuity?

No, and the distinction matters. World models such as Runway's GWM-1 and Google's Genie 3 maintain spatial and physical persistence in real-time explorable environments. Narrative canon (character history, story rules, palette decisions across scenes) is a different problem that they are not designed to address.

What is the fastest way to improve character consistency right now?

Build a multi-view reference sheet on a single canvas rather than using one portrait, and storyboard in 9-panel grids rather than generating panels separately. Those two habits together eliminate the majority of visible drift in most projects.

What does the AI Showrunner do?

It is Magic Lantern's continuity layer. It holds the canon of a living world (characters, locations, palette, visual rules) and enforces it on every scene generated inside that world, across sessions and episodes, so filmmakers do not rebuild continuity from scratch each time.

Your models will keep getting better at frames. They will not start remembering your world on their own.

Magic Lantern is an AI cinematic storytelling platform where filmmakers build living worlds that hold their canon. The Collective is in Public Beta.

Build yours at magiclantern.io.