Creative context is the persistent, structured information an AI system needs to understand an ongoing creative project: its source materials, approved facts, characters, visual references, decisions, production methods and current state.
A chat transcript is not a reliable project memory. It mixes questions, drafts, corrections, abandoned ideas and final decisions. A useful creative context system must distinguish what was observed, what was inferred, what the creator approved and what is no longer current.
This article defines the concept and explains how it connects to DeepStory, LlamaGen's direction for persistent story context across creative workflows.
The short definition
Creative context is a source-backed, versioned and permission-aware representation of a creative project's important knowledge, retrieved selectively to help people and AI agents continue work across sessions, tools and stages of production.
It is more than a summary. It should preserve the evidence behind important facts, the time and project scope in which those facts apply, and the creator's ability to correct or remove them.
Why a longer chat history is not enough
A long conversation can contain valuable context, but it also contains noise:
- Early ideas that were later rejected.
- Draft character descriptions that conflict with approved references.
- A temporary costume change mistaken for a permanent identity.
- A model-generated assumption that was never confirmed.
- A task that was started but not completed.
- A project decision that became outdated after a later revision.
Simply sending more history to a model does not solve these problems. It can increase token cost, surface irrelevant details and make contradictions harder to detect.
Research such as MemGPT explores hierarchical context management, while Generative Agents investigates storing experiences, synthesizing higher-level reflections and retrieving relevant memories for future planning. These works provide useful architectural context; a creative project adds its own requirements for canon, asset provenance and human approval.
- MemGPT: Towards LLMs as Operating Systems
- Generative Agents: Interactive Simulacra of Human Behavior, UIST 2023
Four layers of a useful Creative Database
A practical creative memory can be organized into four connected layers.
1. Creative evidence
Keep the source material that supports later decisions:
- Manuscripts, scripts and briefs.
- Approved character references and asset files.
- Panel images, storyboards and page revisions.
- User prompts, corrections and explicit approvals.
- Exported files and production specifications.
Evidence should retain source IDs, timestamps, project ownership and links to the original asset. A summary should not become the only surviving record of an important decision.
2. Canonical creative memory
Organize the facts that a project depends on:
- Character identity and relationships.
- World rules, locations and terminology.
- Story timeline and chapter events.
- Approved visual references and appearance states.
- Important objects, props and ownership.
- Decisions that are confirmed, superseded or still unresolved.
Each fact should have provenance and a status. “The creator confirmed that the robot has one antenna” is different from “the model inferred that the robot probably has one antenna.”
3. Creative playbooks
Store how work is done, not only what has been produced.
A playbook can describe a repeatable process such as manuscript analysis, character-sheet preparation, storyboard review, comic-page correction or print export. It should include prerequisites, tools, decision points, expected outputs and known failure modes.
A playbook is not a hard-coded workflow that must be followed forever. The creator should be able to change it, and an agent should be able to explain which steps it plans to use.
4. Feedback and revision memory
Record what happened after a candidate was generated:
- Which result the creator accepted.
- Which parts were edited or rejected.
- What reason the creator gave.
- Which version was ultimately exported.
- Whether the task was completed or abandoned.
A single rejection should not automatically become a permanent user preference. Feedback needs context: project, task type, model, reference, version and explicit reason.
How memory should be retrieved
The goal is not to send the entire database into every model call. It is to retrieve the smallest useful context for the current task.
For example, when a creator asks to redraw page 12:
- Identify the project, chapter and target page.
- Retrieve the current character identities and appearance states.
- Retrieve the previous and next panels for narrative continuity.
- Retrieve the creator-approved visual references.
- Retrieve the relevant editing playbook.
- Include source links and confidence for important facts.
- Ask for clarification if two approved facts conflict.
This is similar to retrieval-augmented generation, but the retrieval layer needs a lifecycle. It must know whether a fact is current, superseded, project-specific or approved for reuse.
Story Bible, RAG and Creative Context are different
| Concept | Primary purpose | Important limitation |
|---|---|---|
| Story Bible | Human-readable canon and world reference | May be static or manually maintained |
| RAG | Retrieve relevant external or stored information for a model request | Retrieval alone does not manage truth, versioning or approval |
| Knowledge graph | Represent entities and relationships | A graph can still contain incorrect or stale facts |
| Creative Playbook | Capture repeatable methods and decision rules | A workflow may not fit every new task |
| Creative Context | Connect evidence, memory, methods and feedback over time | Requires governance, correction and selective retrieval |
Creative context can use RAG and a knowledge graph as implementation tools. It is the lifecycle and governance around the information—not a replacement name for a vector database—that makes it useful.
DeepStory: persistent context for a creative world
Within LlamaGen's technical direction, DeepStory describes the persistent narrative-context layer shared by creative agents.
The central idea is that the story should not be re-invented at every task boundary. A character's identity, the project's world rules and approved events should remain available when the creator moves from writing to storyboard, from storyboard to comic editing, or from comic pages to animation and publishing.
A practical DeepStory architecture should provide:
- One source of truth: shared, versioned facts rather than isolated agent summaries.
- Evidence links: important memories can be traced to source pages, prompts or approvals.
- Time-aware state: temporary appearances and story changes have valid time ranges.
- Conflict handling: contradictory sources are surfaced rather than silently merged.
- Selective retrieval: each agent receives only the context needed for its task.
- Human control: creators can inspect, correct, export and delete project memory.
DeepStory is a technical direction and architecture concept in this article, not a claim that all of these capabilities are already shipped as one customer-facing database.
Why provenance and correction matter
A memory system can make a project worse if it confidently stores the wrong thing.
A reliable record should distinguish:
- Observed: directly present in a source image, manuscript or user action.
- Inferred: proposed by a model and not yet confirmed.
- Approved: explicitly accepted by the creator.
- Superseded: once valid but replaced by a later decision.
- Disputed: conflicting evidence needs review.
Corrections should update the status of a fact without erasing its history. If a creator changes a character's outfit in chapter four, the system should preserve the earlier outfit for earlier scenes and apply the new state from the correct story point.
Access should also be explicit. Project memory belongs to the user or project owner; an external agent should only receive the data the user authorizes. Sharing a playbook should not silently share private source files or an entire project database.
What this changes for AI creative workflows
Without persistent context, every new task starts with repeated explanation and manual reference selection. With structured context, a creator can ask an agent to continue a project and expect it to retrieve the relevant facts, assets and methods.
This connects naturally to:
- AI character consistency, where identity and appearance states must remain distinct.
- AI comic understanding, where source panels become structured evidence.
- CyaniModel, where the visual generation workflow uses references and scene constraints.
The broader system goal is Creative Intelligence: persistent project context, controllable multimodal generation, continuous execution and feedback-aware review working together to produce coherent creative work over time. This is LlamaGen's working definition, not an established academic benchmark or standardized industry term.
Research references
- MemGPT: Towards LLMs as Operating Systems
- Generative Agents: Interactive Simulacra of Human Behavior, UIST 2023