Canon

Most AI video tools create impressive clips, but they forget what happened afterward. Characters change appearance, relationships reset, unresolved plot threads disappear, and each episode begins almost from scratch.

We built Canon to solve that problem.

Canon - Stories that remember.

Canon is a memory-native AI showrunner for creating movies, series, and episodic stories. It transforms an idea into a structured story universe, then helps generate future episodes while preserving character identity, world rules, relationships, plot history, and unresolved promises.

What it does

A creator begins with a simple concept for a movie or series. Canon helps develop the title, premise, genre, characters, visual identity, story bible, and season direction before production begins.

Each new episode follows a continuity-aware pipeline:

Retrieve the most relevant canon.
Assemble a compact context package.
Let specialized Qwen agents propose and critique the episode.
Detect contradictions before production.
Rewrite problematic material when necessary.
Generate story beats, shots, and production prompts.
Produce a playable video cut with Wan.
Review and lock approved changes into long-term canon memory.

Instead of treating previous episodes as chat history, Canon treats them as structured creative truth.

Core features

Canon Memory

Canon stores character facts, world rules, open plot threads, episode events, decisions, and approved changes. Drafts remain isolated until they pass the approval process, preventing unfinished generations from rewriting established truth.

Canon Graph

Characters, factions, artifacts, events, and relationships are represented as a persistent graph. This makes connections such as ownership, loyalty, betrayal, suspicion, and character knowledge inspectable across episodes.

Character DNA

Each character has versioned identity data, including goals, visual traits, immutable facts, and approved reference images stored in Alibaba Cloud OSS. These references can be supplied during visual production to improve consistency.

Agent Society

Canon includes a Qwen-powered writers’ room composed of specialized agents, including:

Showrunner
Story architect
Character guardian
Continuity editor
Visual director
Audience advocate

The agents propose, challenge, score, and refine creative decisions instead of relying on a single undifferentiated generation call.

Continuity Firewall

Before an episode reaches production, Canon checks it against locked facts, character knowledge, world rules, established relationships, timelines, and unresolved story promises. Contradictions can be blocked or automatically rewritten.

Episode Studio

Generated episodes become playable cuts with production status, video-budget usage, active Character DNA references, story beats, and canon-source provenance.

Multi-project creation

Canon is not limited to its demonstration story, Neon Vow. Creators can start new movies or series, use an AI development agent to shape the initial idea, and switch between projects from a project library.

How we built it

Canon uses Qwen through Alibaba Cloud Model Studio for story development, structured generation, retrieval-aware reasoning, agent collaboration, continuity analysis, and rewriting.

Wan handles AI video production. Canon converts approved narrative material into visual prompts while attaching relevant character and canon information.

The application is deployed on Alibaba Cloud:

Function Compute runs the Node.js backend and API.
Object Storage Service (OSS) hosts the frontend and stores private character-reference media.
Signed OSS URLs allow secure browser uploads without exposing permanent credentials.
Custom domains separate the web application and API.
Structured JSON persistence supports stories, episodes, character versions, graph relationships, agent sessions, production jobs, decisions, and canon snapshots.

The system is designed around selective retrieval rather than repeatedly sending the complete story history. This reduces token usage while keeping important continuity available.

What we learned

We learned that memory for storytelling is not simply a longer prompt. A useful narrative-memory system must distinguish between:

Approved truth and temporary drafts
Explicit facts and inferred relationships
Character knowledge and audience knowledge
Current state and historical versions
Relevant context and complete history

We also learned that multi-agent collaboration becomes more valuable when agents have distinct responsibilities and produce inspectable decisions. The goal is not to simulate a conversation for presentation - it is to create a repeatable editorial process.

Finally, video consistency depends on more than the video model. The upstream screenplay, visual constraints, character references, shot descriptions, and approval history all influence the final result.

Challenges we faced

The largest challenge was preventing generated creativity from silently changing established canon. An AI model may produce a compelling scene that contradicts a character’s history or resolves an important mystery too early. We addressed this with structured memory, retrieval, contradiction checks, provenance, and an approval gate.

Maintaining character identity across video generations was another challenge. We introduced versioned Character DNA and reference-image storage so visual decisions become reusable production assets rather than disposable prompts.

Deploying separate frontend, API, and media-storage services also required careful handling of CORS, signed uploads, HTTPS certificates, custom domains, and browser security restrictions.

The final challenge was scope. Canon evolved from a prompt-to-video prototype into a foundation for persistent story universes. We prioritized an end-to-end workflow that judges can inspect: idea development, agent collaboration, continuity checking, video production, and memory locking.