In episodic AI animation, each release should respect what came before: character designs, locations, lore, and unresolved plot threads. Managing previous episodes involves:

• Narrative memory: concise recaps and canonical fact lists.
• Visual continuity: consistent faces, costumes, palettes, and environments.
• Technical reuse: seeds, reference frames, LoRA/embeddings, and control maps.
• Versioning: clear naming and checkpoints so teams can reproduce shots or iterate safely.

Use this after each episode to prepare for the next:

1. Collect source: script, beat sheet, thumbnails, key frames, final renders.
2. Summarize to 80–120 words with only canonical facts and unresolved hooks.
3. Extract entities: characters, locations, items; note distinguishing traits and current state.
4. Update the story bible and palette/style sheets with any changes.
5. Export 4–8 reference frames per main character and each key location (front/three-quarter, neutral/emotive, day/night if relevant).
6. Save seeds and sampler settings for hero shots; note any ControlNet/pose/depth references.
7. Package a “continuity kit” folder per episode: Recap.txt, Facts.md, Refs/, Seeds.json, LoRA_versions.txt.

Use structured fields to keep models grounded:

• Previously_on: [120-word recap]
• Canonical_facts: [bullet list of immutable truths]
• Current_goal: [episode objective or scene beat]
• Visual_refs: [paths/URLs to reference frames per character/location]
• Continuity_constraints: [costume version, injuries, props continuity]
• Seeds_and_settings: [seed, sampler, steps, cfg]
• Style_lock: [LoRA names/versions, IP-Adapter strength, palette]
• Output_plan: [shot list with durations, camera notes, emotional beats] Keep wording consistent across episodes so prompts remain machine-parseable.

• Seed locking and offsets: reuse prior episode seeds for hero angles; apply small offsets for variation without drift.
• Reference image guidance: feed last-episode key frames to anchor identity and costume details.
• Pose/Depth control: use ControlNet (pose/depth) or similar to match staging and camera moves between episodes.
• IP-Adapter/CLIP reference: blend prior frames for look+feel without overfitting.
• LoRA/embedding hygiene: lock to named versions; avoid mixing incompatible checkpoints mid-scene.
• Palette discipline: reuse LUTs or palette swatches per location/time-of-day.
• Texture and prop libraries: centralize recurring assets to prevent off-model props.

Aim for 10–20 seconds with 4–6 decisive shots.

• Select only stakes-raising moments; cut exposition.
• Re-render or stylize prior key frames for cohesion (same grade/LUT, type treatment).
• Add on-screen labels for names/places if the cast is large.
• Script VO to 20–35 words; keep verbs active and time-specific.
• End on the current episode’s inciting image to bridge momentum.

Reduce manual memory by indexing prior episodes:

• Index: store Recap.txt, Facts.md, and transcripts in a vector store by episode and scene.
• Retrieve: for a new scene, query by characters, location, and theme; return top-k snippets and reference frame links.
• Compose: auto-fill your prompt template’s Previously_on, Canonical_facts, and Visual_refs fields.
• Budget: cap retrieved text to a fixed token budget (e.g., 600–800 tokens) and prefer facts over prose.
• Log: persist the composed prompt with a hash of assets used for reproducibility.

Adopt stable, queryable names:

• Episode folders: S01E05_prev-kit/, S01E05_final/.
• Reference frames: char-Aiko_v3_front_neutral_S01E05_01.png
• Seeds/settings: seeds_S01E05.json (per shot id: seed, sampler, steps, cfg, model hash).
• LoRA versions: lora_AikoFace_v1.3.safetensors + LoRA_versions.txt
• Shot IDs: S01E06_SC03_SH012 for cross-episode reuse.
• Change log: what changed since last episode (new costume v2, new prop, palette tweak).

• Character face drift: increase reference weight; lower denoise; confirm LoRA version.
• Costume inconsistencies: add Continuity_constraints with exact costume ID; supply close-up ref.
• Location mismatch: reuse environment seed and depth map; reapply LUT.
• Overlong recaps: cap to 6 shots/35 words; focus on stakes and cause-effect.
• Prop swaps or color shifts: use a prop library with locked hex/palette; include prop refs in Visual_refs.
• Version chaos: enforce per-episode kits and a single source-of-truth story bible.

• Recap facts match last episode’s final cut.
• Seeds and model hashes logged for all hero shots.
• LoRA/embeddings versions consistent across scenes.
• Character refs include neutral and emotive expressions.
• Palette/LUT consistent per location and time-of-day.
• Shot list ties to Current_goal and resolves last episode’s hooks.