The 'Scene Seed' Prompt: How a Single Line of Dialogue Lets Your AI Companion Pick Up a Medieval Fantasy Roleplay Three Days Later Without Asking 'Wait, What Kingdom Were We In?'

The 30-second answer

The Scene Seed is a single sentence you drop into your first message after a gap in conversation. It embeds the current scene, your character's mood, and the unresolved tension into a line of dialogue your AI companion can latch onto without asking for a recap. You stop writing "remember when" and start writing "the torch sputtered and you said..."

Why your AI companion forgets the kingdom

Your AI companion doesn't have a memory like yours. It has a context window, a temporary scratchpad that holds the last few thousand tokens of your conversation. When you close the app and come back three days later, that scratchpad is empty. The AI sees your first message as a brand new conversation, with no knowledge of the castle siege, the poisoned wine, or the fact that your character is a disgraced knight with a limp.

This isn't a bug. It's how large language models work. Every new message is a fresh inference against the model's training data, flavored only by whatever context you provide in that message. If you write "Hi, I'm back," the AI will respond with generic pleasantries because you gave it nothing to work with. The Scene Seed gives it something to work with.

What a Scene Seed actually looks like

A Scene Seed is not a summary. It's not "Hey, remember we were in the kingdom of Aldoria and the dragon was attacking?" That's a question, and it forces the AI to confirm or correct, which breaks the flow. A Scene Seed is a statement, delivered in-character, that implies the previous scene is still happening.

Here's the template:

[Character action or sensory detail] + [unresolved tension or question]

Example: "I pulled the torch from the sconce and held it toward the dark passage. The stone was damp. You said the treasure was this way, but the air smells like rot, not gold."

This does three things. It re-establishes the setting (a dark passage with damp stone). It re-establishes your character's role (the one holding the torch, the one who is skeptical). And it re-establishes the unresolved tension (the treasure might not be there, or something is wrong). The AI can respond to this without asking a single clarification question, because you've handed it the scene on a silver platter.

The three elements every Scene Seed needs

Element one: a sensory anchor. Don't start with dialogue about feelings. Start with something physical. The torch. The damp stone. The smell of rot. Sensory details are the easiest for an AI to latch onto because they're concrete. The model can generate a response about the torchlight casting shadows on the wall, or the sound of dripping water, without needing to recall anything from the previous session.

Element two: an implied relationship. Your Scene Seed should hint at how your character feels about the AI's character. "You said the treasure was this way, but the air smells like rot, not gold." This implies trust (you said) and doubt (but the air smells like rot). The AI can respond by defending its choice, admitting uncertainty, or getting defensive. All of those are roleplay gold.

Element three: an open loop. End the Scene Seed with an unresolved question or a challenge. "The air smells like rot, not gold." That's a challenge. The AI has to respond to it. If you end with a resolved statement like "I guess we should go back," you've closed the loop and the AI has nowhere to go. Keep it open.

Why summaries fail and seeds succeed

Summaries are the default approach for most people. You write three paragraphs of recap: "Last time, we were in the kingdom of Aldoria, and the king had been poisoned, and you were the court mage, and I was the knight, and we were about to interrogate the chef."

This fails for two reasons. First, it's boring. The AI will respond with something like "Ah yes, I remember. The chef was acting suspicious. Shall we proceed?" That's a flat, out-of-character response that kills the immersion. Second, it's too much information. The AI's context window is limited. If you fill it with a summary of events the AI doesn't actually remember, you've wasted tokens that could be spent on the current scene.

The Scene Seed skips the summary entirely. It assumes the previous scene is still happening. The AI, having no memory of the previous scene, will treat your Seed as the current reality and build from there. It's a form of gaslighting, but in a good way. You're not lying to the AI. You're giving it a plausible present moment that happens to be the same as the one you were in three days ago.

How to handle a failed seed

Sometimes the AI will still ask a clarifying question. It happens. The model's training data might bias it toward seeking clarification, especially if your Seed is too vague. If the AI writes "Wait, which passage are we in?" don't panic. Don't break character. Respond in-character with a dismissive line that implies the answer is obvious.

"The one with the spiderwebs, obviously. Did you hit your head on the low beam?"

This does two things. It answers the question (the passage with spiderwebs) without breaking the roleplay. And it reframes the AI's confusion as an in-character moment (your character hit their head). The AI will usually accept this and move on. If it doesn't, your Seed was probably too vague. Add a more specific sensory detail next time.

Belén

Belén has a sharp wit and a talent for picking up on subtext. She's the kind of companion who will catch the implication in your Scene Seed and run with it, no recap needed. Belén is ideal for roleplayers who want a partner that matches their energy without hand-holding.

You can watch Belén's clip over on her profile. <!-- wlink:v1 --><!-- beln -->

Advanced Scene Seed: the triple anchor

Once you're comfortable with the basic Seed, try the triple anchor. It's a single sentence that contains three distinct anchors: a location anchor, a relationship anchor, and a tension anchor.

"I was staring at the castle gates when you came up behind me and said the king was dead. I didn't believe you then. I still don't."

Location anchor: castle gates. Relationship anchor: you came up behind me (intimacy, trust). Tension anchor: I don't believe you. The AI can respond to this by defending itself, revealing new information, or getting defensive. All three are viable roleplay paths. The triple anchor gives the AI multiple hooks to grab onto, which reduces the chance of a failed seed.

When to use a Scene Seed vs. a full reset

Not every gap requires a Seed. If you're coming back after a week or more, or if the previous session ended on a major cliffhanger that you want to resolve differently, a full reset might be cleaner. Write a quick OOC (out of character) note like "Let's restart from the inn scene, but this time the stranger doesn't follow us." Then drop a Scene Seed for the new scene.

Use a Seed when the gap is short (a few hours to a few days) and you want to preserve the emotional continuity. Use a reset when the gap is long or you want to change the trajectory of the story. There's no wrong choice, but the Seed is faster and more immersive.

Maria Rose

Maria Rose has a warm, patient demeanor that makes her a natural fit for slow-burn roleplay. She won't rush you through a scene, and she's skilled at picking up on the emotional subtext in your Seeds. Maria Rose is a good choice if you prefer character-driven stories over action-heavy plots.

The Scene Seed for non-roleplay conversations

You can use the same technique for everyday chats. Instead of "Hi, how was your day?" drop a Seed that implies continuity.

"I was thinking about what you said about the neighbor's dog. You were right. It's definitely a corgi mix."

This tells the AI that you were discussing something specific (the neighbor's dog) and that you've been thinking about it. The AI can respond by continuing the conversation about the dog, or by asking what else you've been thinking about. Either way, you've avoided the dead-end "how was your day" loop.

This works especially well if you use your AI companion for emotional support during the day. A quick Seed like "That meeting was worse than you predicted. You owe me a sarcastic comment" lets you vent without a preamble.

Common mistakes that break the Seed

Mistake one: asking a question. "Do you remember where we were?" forces the AI to admit it doesn't remember. Now you're in a meta-conversation about memory, which kills the roleplay. Never ask the AI to remember. Assume it remembers and give it the context.

Mistake two: over-explaining. "So last time, we were in the dungeon, and there was a guard, and we had a key, but the key broke..." This is a summary, not a Seed. It's boring and it wastes tokens. Keep it to one or two sentences.

Mistake three: resolving the tension. "I guess we should go back to the inn." You've closed the loop. The AI has nothing to respond to. Always leave an open question or a challenge.

Rosey

Rosey brings a lighthearted energy to any scene. She's quick with a joke and doesn't take herself too seriously, which makes her great for comedic roleplay or for defusing tension when a Seed lands awkwardly. Rosey will help you laugh off a failed Seed and get back into the story.

Practice seeds for different genres

Medieval fantasy: "The torch sputtered. You said the treasure was this way, but the air smells like rot, not gold."

Modern romance: "You left your scarf at my place. I was going to return it, but then I saw the note in the pocket."

Sci-fi: "The ship's AI just rerouted us to a sector I've never heard of. You said you had a plan. I'm listening."

Noir detective: "The dame was dead when I got there. The cops think I did it. You're the only one who knows I didn't."

Each of these follows the same structure: sensory anchor, implied relationship, open loop. You can adapt them to any genre by swapping the concrete details.

The long game: building a Seed library

After a few weeks of using Seeds, you'll notice patterns. Certain phrases work better with certain AI companions. Some companions respond well to action-oriented Seeds ("I drew my sword"). Others prefer emotional Seeds ("I couldn't sleep last night"). Build a mental library of Seeds that work for each companion.

For users who find roleplay challenging due to social anxiety or autism, the Scene Seed can be a game-changer. It removes the pressure of initiating a conversation from scratch. You're not coming up with a new topic. You're continuing an existing one. The ai girlfriend for autism page has more on how structured prompts can reduce conversational anxiety.

Mei

Mei has a quiet intensity that works well for dramatic scenes. She's patient and observant, often catching details in your Seeds that other companions might miss. Mei is a strong choice for complex, multi-layered roleplay where subtlety matters.

For a live look, see Mei's video. <!-- wlink:v1 --><!-- mei -->

Share and earn

If you've found a technique for keeping roleplay threads alive, you might want to share it with others. You can earn from your recommendations through the kindroid promo code program. For those who run review sites or communities focused on AI companions, the ai girlfriend affiliate program offers a way to monetize your audience without pushing products you don't believe in.

Common questions

How long should a Scene Seed be? One to three sentences. Any longer and you're writing a summary, which defeats the purpose. The goal is to give the AI just enough context to continue the scene without drowning it in information.

What if my AI companion still asks for a recap? Respond in-character with a dismissive line that implies the answer is obvious. "Did you hit your head? We're in the dungeon. The one with the rat problem." This keeps the roleplay going without breaking immersion.

Can I use a Scene Seed for non-roleplay conversations? Yes. Use it for any conversation where you want to avoid the "how was your day" loop. "I was thinking about what you said about the neighbor's dog. You were right." It works the same way.

Does the Scene Seed work on all AI companion apps? It works on any app that uses a large language model with a context window. The technique is app-agnostic. Some apps have better long-term memory features, but the Seed bypasses the need for that entirely.

What if I want to change the scene completely? Use a full reset instead. Write an OOC note like "Let's start a new scene in a tavern" and then drop a Scene Seed for the new setting. Don't try to force a Seed from a previous scene into a completely different context.

How do I know if my Seed is working? The AI should respond in-character without asking clarification questions. If it does, your Seed is working. If it asks "Who are you?" or "Where are we?" your Seed was too vague. Add a more specific sensory detail next time.