A personal agent memory architecture is a permissioned system for storing, reconciling and retrieving the information an AI agent needs across tasks and sessions. It must distinguish observed evidence, authoritative facts, working state and learned procedures. In creative production, it also needs to preserve the boundary between a fictional story's canon and the human creator's real preferences.
Quick answer: A good memory system remembers what was decided, why it was decided, which version is current and how the next agent may use it. A single giant prompt, character-description paragraph or transcript archive is not enough.
Four layers of a creative personal database
| Layer | Typical record | Role in the next task |
|---|---|---|
| Evidence | Original manuscript, approved image, edit event, timestamp and source | What actually happened |
| Canonical story memory | Character identity, relationships, timeline, world rules, current scene state | What must remain coherent |
| Procedural playbook | Storyboard review order, production preferences, publisher checks | How this creator usually works |
| Feedback and outcomes | Accepted/rejected draft, localized repair, published export | Whether the method delivered |
A semantic search index can support retrieval, but the authoritative record should live in a structured, versioned source that the creator can inspect and correct. Embeddings and summaries are useful derived representations; they must not silently replace the original evidence.
The data model: identity is not scene state
Consider a recurring character. Some attributes define their durable identity, such as face structure, distinctive hairstyle and name. Others change legitimately between scenes: outfit, injury, location, relationship status or held props.
An implementable design separates at least these records:
| Record | Example | Update rule |
|---|---|---|
| Character identity | Botanist Mira, short dark curls, distinctive face | Change only with explicit approval and a source |
| Appearance state | Teal jacket removed after the greenhouse scene | Version by chapter, time and scene |
| Reference asset | Approved front/side character sheet | Preserve immutable asset ID and revision |
| Story event | Mira leaves the scarf on the bench | Link to manuscript passage and approved storyboard |
| Production rule | Keep speech balloons editable | Treat as project/playbook constraint, not character canon |
| User preference | Prefers restrained camera moves | Infer cautiously; allow correction and expiry |
Without this separation, a model might incorrectly "fix" a deliberate costume change or carry a temporary prop into an unrelated chapter.
The agent memory lifecycle
Capture with source fidelity
A memory item records its source type, project scope, event time, content hash or stable asset ID, and whether the creator explicitly approved it. Conversation text and generated imagery should not become canon merely because they exist.
Extract candidates, not conclusions
An AI model may suggest that two references depict the same character or that a certain framing preference has recurred. Those suggestions should remain provisional until evidence or confirmation supports them. When ambiguity matters, the agent should ask rather than silently promote a guess to fact.
Resolve conflicts and superseded details
A new canon revision can supersede an earlier detail, while a separate draft or alternate universe remains valid within its own branch. Corrections must be recorded with reason and scope; deletion should make the invalid fact unavailable to later retrieval.
Retrieve by task intent
For a close-up, facial references may matter more than a full-body pose. For a long shot, silhouette and costume shape can dominate. For dialogue lettering, source text and speaker attribution may matter more than pixel similarity. Retrieve the smallest context packet that answers the current task.
Verify outcomes and update selectively
Compare the generated artifact against the requested scene, approved identities, correct cast count, source dialogue and intended output format. Store the decision and failure class, not merely a model's enthusiastic self-rating.
An example context packet for a comic panel
| Field | Example content | Provenance |
|---|---|---|
| Project and scene | Greenhouse, chapter 4, panel 2 | Storyboard revision |
| Required cast | Mira and one mint robot | Approved scene plan |
| Stable identity | Mira's approved reference sheet | Character asset ID |
| Current state | No coral scarf in this scene | Accepted prior panel |
| Action | Robot lifts the terracotta pot | Manuscript passage |
| Camera | Medium shot, right-to-left interaction | Storyboard decision |
| Forbidden drift | Do not add a third character or text | User-approved brief |
This is a proposed architecture illustration. It does not imply that LlamaGen already exposes a universal persistent context-packet API.
Memory governance is part of the technical design
A personal database should answer who owns the data, which project or agent can read it, how long it persists and what happens when someone corrects or deletes it. A cross-project search must not turn a private client's character reference into another customer's prompt.
Minimum controls: per-project authorization, explicit sharing grants, provenance, immutable revisions, revocation, selective deletion, user-visible corrections and export. A user-controlled collection of context is only portable when it can be exported in a documented, usable format without hidden dependencies.
DeepStory, CyaniModel and the memory boundary
In LlamaGen's architectural direction, DeepStory would own persistent narrative context and canonical facts. CyaniModel would use the relevant visual references and constraints to support creation or inspection. An Agent Loop would carry task state through generation and review, while feedback learning would distinguish accepted panels from discarded attempts.
Do not collapse these into one catch-all term. Visual generation is not the database; task orchestration is not the visual model; and retrieving a preferred style is not evidence that the result improved.
For the visual side, see consistent character comics and comic storyboard generation. For the broader agent definition, see What Is a Personal AI Agent?.
The memory evaluation protocol
A serious test should use the same fixed story, character references and source excerpts for all compared approaches. Compare a baseline with no memory, a long-context transcript baseline, a summary-only baseline, and a versioned evidence-backed memory baseline.
Measure:
- Fact freshness: Did the system avoid superseded wardrobe and relationship states?
- Attribution: Can the correct source be retrieved for each consequential claim?
- Identity continuity: Do the same characters remain recognizable across perspective changes?
- Error containment: Does a local repair preserve already accepted neighboring content?
- Creator effort: How much work was needed to reach the accepted page?
- Privacy boundary: Can the system prevent unauthorized cross-project context retrieval?
Do not report a percentage advantage without actual test inputs, sample counts, evaluation rules and failures. A successful single demonstration is not a benchmark.
Is vector search enough for personal AI memory?
No. Vector search can find semantically similar material; it does not itself establish which fact is approved, current, or authorized for this project.
Should an agent remember every preference forever?
No. Some preferences are temporary, some are project-specific and some may be inferred incorrectly. Attach evidence, timestamps, confidence and correction controls to candidate preferences.
What is the relationship between story bible and personal database?
A story bible contains the fictional world's authorized facts. A personal database may also contain the creator's methods, preferences and project actions. These scopes must not be conflated.
Can I share a playbook without sharing my original manuscript?
That should be a design requirement of a permissioned system. Sharing reusable procedures and exposing private source materials are different permissions; this page does not claim that cross-agent sharing is already released.
Research references
- IBM Research: Personal Agents and Conversational Memory
- ACL Findings 2026: Memora and forgetting-aware memory
- Microsoft multi-agent memory reference
- ViStoryBench identity and visual-story metric definitions
- LlamaGen's documented CyaniModel research boundaries
Technical editorial framework, November 6, 2026. Architecture examples are proposals. No deployed personal-database product, independent Cyani foundation weights or measured performance gain is asserted.