When AI Worlds Learn to Remember
The next creative frontier is a place that stays itself when the camera turns away, returns and asks to see more.
Page 1
BUNCOSMOS | CREATIVE AI
CREATIVE AI | BUNCOSMOS DEEP DIVE
When AI Worlds Learn to Remember
The next creative frontier is a place that stays itself when the camera turns away, returns and asks to see more.
22 October 2026 | BunCosmos | 11 minute read
Imagine entering Willow Hollow on an autumn afternoon. There is a little house
beside the path, a lantern near its door and a crooked tree leaning over the garden
wall. You walk past, turn a corner and look back. The lantern is still there. The house
has the same windows. The tree has kept its peculiar shape.
Nothing dramatic has happened, yet a demanding creative promise has been kept. The place
has remained itself. For an artist building a fictional world, that quiet continuity can matter
as much as the beauty of any single frame. It allows a setting to gather meaning. A window
can become familiar. A doorway can mark the beginning of a story. An audience can
recognize a place before anyone says its name.
This article uses that imagined Willow Hollow walk as a creative example, not as a claim that
a current tool can reproduce the project faithfully. It points toward an active research
question: how can generated environments preserve their identity while people move
through them? Recent work on visual memory and three-dimensional consistency is making
that question increasingly concrete.
A beautiful view is only the beginning
A single picture can suggest an entire town while showing only one side of a street. To move
a camera through that town, a system must make further decisions. What is behind the wall?
How wide is the path? Where does the roof end? If the camera returns, do the answers
remain compatible with what was shown before?
In an ordinary creative workflow, an artist might answer those questions with plans, models
and continuity notes. A generative workflow also needs some way to constrain its future
choices. Otherwise, the hidden side of a building can be invented one way on the outward
journey and another way on the return. The difficulty is not simply producing more visual
detail. It is preserving the commitments already made.
The term world model covers different technical approaches, so it helps to ask what a
particular system produces. Some generate new video as controls arrive. Some build an
explicit spatial representation that a renderer can revisit. Others infer new camera views
from a small set of images. These can support overlapping creative goals while offering
different kinds of control and persistence. The distinction is visible across the research and
products discussed below.
buncosmos.ai | 22 October 2026 1
Page 2
BUNCOSMOS | CREATIVE AI
When the picture becomes a place to explore
Google DeepMind introduced Genie 3 in August 2025 as a model generating interactive
environments from text, reporting 720p output at 24 frames per second and consistency
over a few minutes. It described visual memory extending roughly a minute into the past. Its
limitations included restricted actions, imperfect geographic accuracy and difficulties with
complex interactions among independent agents. Those are model-level claims from the
original announcement, rather than guarantees of unlimited simulation. [3]
Google’s January 2026 Project Genie announcement described a user-facing experimental
prototype powered by Genie 3, with world creation, exploration and remixing. That release
limited generation to 60 seconds and omitted some previously demonstrated model
capabilities. The difference is worth preserving: a research model’s capabilities and the
limits of a particular product experience are separate facts. Launch access also had
subscription, age and geographic restrictions; the original announcement should not be
treated as a current worldwide availability guide. [4]
For storytelling, real-time navigation changes the relationship between viewer and image. A
filmmaker usually decides where the audience looks next. An explorable environment allows
the visitor to make some of those choices. That possibility brings a different artistic demand.
The setting must support a curious person who might look behind the scenery, pause at an
unimportant doorway or return to a place the author expected them to leave.
One route to continuity is to build the space
World Labs’ Marble follows an explicit 3D route. Its November 2025 announcement
describes worlds generated from text, images, video or rough 3D layouts, with tools for
editing, expanding and combining them. It also describes export as Gaussian splats, meshes
or video. A stored spatial asset offers a different starting point from repeatedly generating
the next image. It can become material for a downstream creative workflow. Those
capabilities do not, by themselves, establish a complete game with working characters, rules
and physically accurate interactions. [5]
Two earlier graphics advances help explain the vocabulary. NeRF, introduced in 2020,
represents a scene through a neural function learned from images with known camera
positions. New views are rendered by querying that representation. It demonstrated that a
scene’s appearance could be represented in a form that supports views beyond the original
photographs. Reconstructing a scene this way is different from freely inventing a complete
fictional environment, but it helped make learned scene representations a practical subject
of visual computing. [6]
The 2023 3D Gaussian Splatting paper uses a different representation: many spatial
Gaussian elements whose properties are optimized to reproduce a scene’s appearance. Its
renderer made high-quality novel views possible at real-time rates in the reported
experiments. A useful visual analogy is a cloud of soft, oriented marks arranged in space.
These are not simply ordinary painted pixels, and a splat representation does not
automatically include collision rules, physical behavior or a rigged character. [7]
buncosmos.ai | 22 October 2026 2
Page 3
BUNCOSMOS | CREATIVE AI
For a creator, the practical question is what remains editable after generation. A beautiful
rendered clip may be enough for one purpose. Another project may need the camera to move
again tomorrow, from a position that was never requested today. A third may need a
doorway widened while the surrounding architecture remains intact. These are different
deliverables, and choosing among them is part of directing the work.
Another route is to remember what the camera saw
WorldMem, first submitted in 2025, introduces a memory bank containing earlier frames
together with state information such as camera poses and timestamps. Its memory attention
retrieves relevant past observations so a generated scene can remain consistent after large
changes of viewpoint or time. Timestamps also help represent how a scene changes, rather
than assuming everything is permanently static. The published framework is a research
approach with reported experimental results, not a promise that every future scene will be
remembered perfectly. [8]
Video World Models with Long-term Spatial Memory, another 2025 study, stores information
in a geometry-grounded memory to help retain environments across revisits. Its authors
report improvements in quality and consistency relative to their baselines. Both projects
make a useful technical distinction visible: information that is no longer in the model’s short
active history can still be preserved elsewhere and retrieved when needed. Their
mechanisms and evaluations differ, so they should not be treated as interchangeable
implementations. [9]
An analogy is a filmmaker’s continuity archive. The current shot cannot contain every earlier
shot, but a record can preserve the position of a chair, the shape of a room and the state of a
door. In a generative system, the archive must do more than exist. It has to retrieve the right
information for the current viewpoint and use it without introducing contradictions. This is a
computational design problem; the word memory here does not imply human recollection or
subjective experience.
World-T3 turns remembered views into a compact learned state
The October 2026 World-T3 preprint investigates a different memory design. It writes past
observations into dedicated, updateable model parameters and retrieves scene features
using camera pose, meaning position and orientation. These updates affect a dedicated
memory state; they should not be confused with fully retraining the foundation model during
every camera move. The memory has fixed-size storage rather than an indefinitely
expanding collection of past frames. In its long-horizon revisit test, the authors report a
PSNR increase from 13.52 to 18.31 decibels, an image-similarity measure for which higher
is better against their strongest baseline, alongside reduced camera rotation error. They
report 20.6 frames per second at 1280 by 704 on one NVIDIA B200 GPU. These are
benchmark and hardware-specific results. Fixed memory size does not mean unlimited
capacity or perfect recall. The authors acknowledge camera-estimation errors, drift and
visual degradation during extended generation. World-T3 is a promising research
demonstration of compact scene memory, with limitations that remain relevant to
production use. [1]
buncosmos.ai | 22 October 2026 3
Page 4
BUNCOSMOS | CREATIVE AI
The creative interpretation is simple: a system should have a better reason for drawing the
same doorway again than the general likelihood that a doorway belongs in that scene. Some
information about this particular doorway should constrain the result. The research
approaches differ in how they preserve that constraint. For an artist, its value would be
visible in less time spent repairing continuity and more freedom to explore a setting.
How to test whether a world stays itself
Orbia, an October 2026 technical report, evaluates 27 models using 720 cases and six
exploration-and-revisit protocols. It separates three questions. Input preservation asks
whether the supplied scene survives. Generated-content persistence asks whether newly
invented regions survive a return visit. Three-dimensional self-consistency asks whether
different views can belong to the same coherent space. The authors find that strong visual
quality and accurate control do not guarantee 3D consistency. Minute-long exploration
exposes forgetting and geometric deterioration that attractive frames can conceal. Its tests
combine appearance and geometry measurements. A system can preserve the supplied
image while forgetting a region it invented later. Orbia is an evaluation framework, not a
model that repairs those failures. Its importance is the separation of capabilities that a
single appealing demonstration can blur. [2]
For a fictional village, those questions translate naturally. Does the original cottage retain
its shape? Does the alley invented after the first turn remain in the same place? Do the views
from opposite ends of the street fit together? These examples are an editorial way to inspect
a creative result, not a substitute for the benchmark’s measurements. They encourage a
habit that is valuable long before a technology becomes reliable: return to the scene of the
promise and see whether it was kept.
Camera control is its own creative instrument
Stability AI’s Stable Virtual Camera research announcement, released in March 2025,
describes generating new views from one or more input images along specified camera
trajectories. It supports paths such as orbits and dolly zooms, aiming for consistent view
synthesis without requiring a separately reconstructed 3D representation. The release was
offered for research under a non-commercial license. It illustrates both the appeal of
directing camera movement and the need to distinguish a research demonstration from a
tool licensed for a commercial production. The linked release documents that status at
introduction. [10]
Camera movement carries meaning. A slow approach can make a house feel welcoming. A
low orbit can turn a familiar object into something monumental. Returning to a previous
position can connect one scene to another. The artistic benefit of consistency is that these
choices can remain intentional. A camera move should change the viewpoint because the
director chose it, while the setting retains the features the story depends on.
buncosmos.ai | 22 October 2026 4
Page 5
BUNCOSMOS | CREATIVE AI
What a persistent world could give independent creators
For an independent artist, the most useful future may be a reusable setting that gradually
acquires detail. A town square could host several episodes. A room could appear in different
seasons. A café could become familiar through the repeated placement of its windows,
counter and tables. This is a creative possibility suggested by the direction of the work,
rather than an assertion that current systems reliably deliver the complete workflow.
Such a workflow would also need controlled change. Continuity does not mean freezing
everything. A chair may move because someone used it. A shop may close for winter. An
artist may decide that the garden wall was always too high and revise it. The system would
need to preserve the difference between an intended edit, a story event and an accidental
change introduced during generation.
Character identity adds another requirement. Remembering a room does not establish that a
model can preserve a particular face, hairstyle, body proportion or outfit across a film. A
creator might need separate character assets, reference views, animation tools and careful
review. A coherent environment could be an important part of that production, while
remaining only one part. This distinction is particularly relevant to projects whose
characters are already recognizable and should retain their original design.
For Willow Hollow, the inviting possibility is therefore broader than one spectacular
fly-through. It is a setting whose accumulated decisions remain available to its creator. An
artist could return to a corner that has already appeared in a story and direct a different
scene there. The audience’s recognition would come from continuity that the artist has
chosen, rather than from an approximate resemblance recreated by chance.
Persistence is also a question of authorship
A creative world is made of decisions about what belongs. If every new view silently revises
those decisions, the artist has to keep reasserting them. A more dependable environment
could move some of that effort into developing the world itself: deciding how its streets
connect, which objects carry a history and what a visitor should notice first.
That does not remove the pleasure of surprise. Generative systems can suggest an
unexpected composition or a corner of the town the artist had never considered. The
important creative freedom is being able to accept that discovery deliberately. Once
accepted, it should become part of the work that future generations respect. The artist
needs both the possibility of invention and the ability to say: keep this.
The frontier is therefore more demanding than endless novelty. It asks whether a generated
place can support attachment, revision and return. A lantern beside a doorway may seem a
small detail compared with the spectacle of a newly imagined world. But when the camera
comes back and the lantern is still where the artist placed it, the world begins to offer
something a passing image cannot: the possibility of becoming familiar.
buncosmos.ai | 22 October 2026 5
Page 6
BUNCOSMOS | CREATIVE AI
Sources and further reading
Research checked on 22 October 2026. Numbered links lead to primary sources. Recent preprints and technical reports describe their authors’ findings and may change with further review. Creative examples and future possibilities are identified as interpretation rather than demonstrated product capabilities.
[1] World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models Wu and colleagues. Preprint, 12 October 2026. Main results and Appendix F limitations.
[2] Orbia: Do Generated Videos Form a Consistent 3D World? Lu and colleagues. Technical report first submitted 14 October 2026; version 2, 18 October 2026.
[3] Genie 3: A new frontier for world models Google DeepMind. 5 August 2025. Model capabilities and limitations as originally described.
[4] Project Genie: Experimenting with infinite, interactive worlds Google. 29 January 2026. Prototype launch and its distinct product limitations.
[5] Marble: A Multimodal World Model World Labs. 12 November 2025. Persistent 3D generation, editing and export.
[6] NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis Mildenhall and colleagues. ECCV 2020. Official project page and method.
[7] 3D Gaussian Splatting for Real-Time Radiance Field Rendering Kerbl and colleagues. ACM Transactions on Graphics, 2023.
[8] WorldMem: Long-term Consistent World Simulation with Memory Xiao and colleagues. First submitted April 2025; version 3, 1 January 2026.
[9] Video World Models with Long-term Spatial Memory Wu and colleagues. Preprint, 5 June 2025.
[10] Introducing Stable Virtual Camera: Multi-View Video Generation with 3D Camera Control Stability AI. Research announcement, 18 March 2025. Capabilities and research-use licensing at release.
buncosmos.ai | 22 October 2026 6