How Stage Whisper Handles Narrative Consistency on Long Campaigns

ai stage-whisper memory ttrpg

The first time an AI quoted something I said three sessions ago, I knew this was different.

Not "approximate" recollection. Not improvisation that sounded like it might have happened. An exact quote, delivered at exactly the right moment, by an NPC who wasn't even the focus of the scene.

That's what Stage Whisper's memory system makes possible. But to understand why it matters, you first need to understand why most AI roleplay tools fail at this.

The Problem: Context Isn't Just Limited, It Gets Diluted

AI models have a limited context window. Aside from the embedded knowledge the AI has from its training, the only thing it knows about your campaign comes from what's in the context. A good modern LLM has a context of around 200k tokens. Larger LLMs have 1 million, but tend to be expensive to run. (This is as of Q1 2026 anyway, I'm sure this post will get outdated very quickly given advances)

200k tokens can store the dialog from roughly one to three sessions of TTRPG gameplay. But just having the full session dialog in the context isn't a foolproof way to ensure it actually gets used. There's also the factor of: context dilution.

AI models have a limited ability to find and extract relevant information from whatever's in the context. The more stuff crammed in there, the thinner the AI's attention is spread. If the context is full of multiple sessions of gameplay, the AI might fail to find the one specific thing the player character is trying to reference, or the thing the narrative is trying to call back to.

The result: inaccurate details, narrative inconsistencies, characters who sound confused. For example, the narrative might write about "the tender moment you shared in that tavern", and it might get the location wrong, or the wrong character, or invent details that never happened. These glaring inconsistencies break immersion fast.

So context dilution can be as damaging as not having the information in the context at all. Both lead to the same problem: characters (or the narrator) not knowing things they should know.

The Four-Layer Memory System

Stage Whisper solves this with four distinct types of memory, each solving a different problem.

Layer 1: Character Sheets

The first problem Stage Whisper solves is just the context not being big enough to hold the entire history of the game. We need to remember who the characters are from session to session. We can't have every new session introduce a whole new cast.

Stage Whisper writes out full character sheets for every major character, including the player character. That information is stored in the database and loaded into context for any scene the character is present in. So the AI knows, at all times, who these characters are: their personalities, their appearance, their backstories.

It doesn't have to load every NPC! So we can have a very big and diverse cast of NPCs that can be brought in as needed. I've run a couple campaigns so far where I'm part of a crew, of a large ship, or a train, and Stage Whisper is able to maintain that crew roster of about 20 characters that can appear, with consistent personality, where needed. The doctor in the medical bay, chief engineer and the engineering crew in the engineering section, cargo chief in the cargo bay, pilot and captain on the bridge, and so on.

Layer 2: Session Recaps

Instead of keeping the whole chat history in context, Stage Whisper summarizes sessions into recaps.

Only the last session recap is loaded into context, along with a list of the five previous session summaries (even shorter one-paragraph version of the recap). The AI can choose to get more session summaries, and load the full recap for any session it decides it really needs to reference. It's not elegant, but it ensures important information is either at the start or the end of the context, where the AI is best able to focus on it.

Between character sheets and session recaps, the AI knows who the characters are and has a good sense of what happened recently. This is in addition to the campaign files which track long-running threads, and are effectively the GM's notes on what's going on in the campaign and what the most recent pressing matters are

Layer 3: NPC Memories and Scene Summaries

But that's often not enough. In longer campaigns, the player and other characters have a history together. They've experienced high-stakes and high-emotion events. They've shared running jokes. They want to reference previously successful gambits, or recall failures.

Session recaps give you the big picture: plot beats, major character development moments. But you need something more detailed for the nuance of specific events or specific quotes.

Stage Whisper also generates NPC memories and scene summaries. The difference: session recaps are focused on the objective narrative of what happened. NPC memories focus on what NPCs were feeling or thinking at the time, their version of events, which may be different from the dry summary. Scene summaries capture specific dialog and action sequences in more detail than the recap.

At the end of sessions, the AI reads through the message history and extracts the important memories for each NPC. It rephrases these into a standardized format containing:

The AI then calls a tool to insert this into a vector database, where the content is vectorized. This provides semantic search, allowing memories to be match by meaning rather than just ekywords.

Later, when the AI needs to search for memories, it writes a query like "Yuki crying in the hallway," and this brings up all relevant memories even if they don't match the exact phrasing. The AI picks what it needs from the results.

Layer 4: Lore

Characters need to remember what happened. But the world itself also needs to stay consistent. A city's political structure needs to be consistent across sessions. A magic system that works a specific way needs to stay consistent too.

Most campaigns don't have all of this planned out from the start. So lore entries get added and updated as sessions run, and then serve as reference for new scenes to ensure consistency.

What It Feels Like When It Works

The distinguishing feature between playing a game and reading a book is that the player has agency. You decide on your actions, your words, and the narrative and other characters respond to that. The fact that characters can quote what you said several sessions ago helps underscore that agency and the connection you're building with the world and the people in it.

Here's a real example from a Stage Whisper session:

[Player input] "I'm glad I was hungry and decided to open my door at 2am"

[Narrative]

She sets the mug down. Slowly. Carefully. Like she's handling something fragile. Then she turns to face you on the couch, tucking one leg underneath her, her knee almost brushing yours.

"You opened your door at 2am because you were hungry."

"Yes."

"And I was crying in the hallway because I was... not okay. And you saw me. And you said, 'I'm not peeping.'"

The context: the NPC and my PC are reminiscing about the night they first met. The NPC has just admitted she was glad to be there because she got to meet my PC. I reciprocate by recalling a specific detail from that night. The NPC responds with a direct quote of what I had said a few sessions back, which could only have come from NPC memory or scene memory.

This scene hits harder because the NPC is able to reference something I said, word for word. Not just improv or make it up, or recall the dry events. It's exact recall. And that specificity makes the relationship feel real. My actions and words mattered.

What Stage Whisper Doesn't Solve Yet

Even with this memory system, there are still problems.

Within-session context: The memory system is effective session-to-session, but doesn't solve the context issue within a single large session. Sessions can run long and contain a lot of dialog. Optimization for context management within sessions is still needed, to avoid that context dilution effect.

Secrets: Because every NPC sheet is loaded and the AI has full view of all characters, it tends to ignore when a particular character shouldn't know or act on something. AIs tend to act on all information available to them, so it's difficult to get them to not act on something a character wouldn't know. That needs a more elaborate setup.

Both of these are active areas of work. The secrets problem in particular is something I'm tackling right now — more on that in a future post.

Stage Whisper is built for people who want to play solo TTRPGs with the depth and consistency of a long-running campaign. Get started or join our Discord to see how the memory system works in practice.

For more on Stage Whisper's memory system, see Part 2: Context Isolation and Part 3: NPC Dreaming.