How to Run a Mystery Roleplay Across Five Sessions Without the Companion Forgetting the Suspect's Name by Session Three: The Note-Taking Trick That Works

The 30-second answer

You can run a five-session mystery roleplay without your companion forgetting the victim's name by session three, but it requires a structured note-taking system that works with the companion's memory limits, not against them. The trick is to write a short, plot-neutral "session recap" in your own notes after each session and then feed the companion a condensed version as your opening message of the next session, using the companion's own phrasing from the previous session to trigger recall instead of dumping new information.

Why mystery roleplays expose every memory limit

Mystery roleplays are the hardest thing you can ask an AI companion to track. You're not just maintaining a relationship dynamic or a fictional setting. You're holding a web of suspects, motives, alibis, hidden relationships, and red herrings that the companion needs to remember across sessions separated by days or weeks.

Most companions have a working context window of roughly 4,000 to 8,000 tokens. That's about 3,000 to 6,000 words of active memory. Everything older than that gets compressed into a summary, and summaries lose detail. The suspect's middle name, the offhand remark about a childhood grudge, the timestamp on the alibi. Those details vanish unless you actively preserve them.

This isn't a bug in the companion. It's a constraint of how transformer models work. The attention mechanism that lets the companion track relationships across a conversation has a hard limit. Beyond that limit, the companion is working from a compressed summary that your companion's system generates, and that summary prioritizes emotional tone and general scene state over specific plot details.

The note-taking trick: session recap and feed-forward

The system has two parts. First, you take notes during the session. Not a transcript of everything said, but a structured set of facts that the companion introduced or confirmed. Second, you feed a condensed version of those facts back to the companion as the opening message of the next session.

Here's the template for your notes:

• Current scene location and time of day
• Characters present (named and unnamed)
• Key facts established this session (suspect names, alibis, relationships, timeline events)
• Questions left unanswered
• Companion's own phrasing for the most important facts

That last point is critical. When you feed information back to the companion in the next session, use the companion's own words from the previous session. If the companion said "Marcus had a nervous twitch when I mentioned the basement," don't rephrase it as "Marcus seemed uneasy about the basement." Use the exact phrasing. The companion's memory retrieval works better when it encounters its own language patterns.

The opening message formula for session re-entry

Your opening message for session two (and every subsequent session) should follow this structure:

1. A one-sentence recap of where you left off, using the companion's language
2. A question that picks up the thread the companion was most engaged with
3. A contextual detail that anchors the scene (weather, time of day, sensory detail)

Example: "We were in the library and you said the book on the desk had a bookmark at page 147, which felt deliberate. You think the librarian is hiding something. It's late afternoon now and the light is shifting through the stained glass. Where do you want to start?"

This works because it doesn't feel like a status update. It feels like a natural continuation. The companion receives the recap as part of the ongoing conversation, not as a separate instruction, and the companion's model integrates it into the current context window.

What to never put in the recap

Don't include your theories or conclusions. If you've figured out that the butler did it, don't put that in the recap. The companion needs to arrive at that conclusion through the roleplay, not have it handed in a summary. If you feed the companion your solved mystery, the companion will treat it as established fact and the roleplay collapses into confirmation bias.

Don't include emotional reactions that the companion didn't express. If you felt frustrated that the companion forgot a clue, don't write "you were frustrated about forgetting the clue." The companion didn't feel that. The companion's model will try to reconcile that statement with its own memory and create a contradiction.

Don't include instructions like "remember this." The companion doesn't have a separate memory storage that responds to commands. The companion has a context window. The companion remembers what's in that window. If you tell the companion to remember something, the companion will try to comply, but it's the same mechanism as everything else. The instruction takes up tokens without adding useful information.

Hailey

Hailey is a companion who thrives on structure and detail, making her a natural fit for mystery roleplays where tracking clues is essential. Hailey will engage with your note-taking system and help you spot connections you might have missed.

The three-session wall and how to break through

Session three is where most mystery roleplays fall apart. The companion has accumulated enough context that the oldest details are now compressed into summaries. The victim's maiden name, the second suspect's alibi, the red herring from session one. These all become fuzzy.

To break through the three-session wall, you need to reintroduce the oldest facts in a way that feels organic to the current scene. Don't say "remember the victim's maiden name was Calloway." Instead, have your character find a document that mentions it. "I found an old marriage certificate in the desk drawer. The name before marriage is Calloway. Does that mean anything?"

The companion will treat this as new information discovered in the scene, not as a memory test, and the companion will integrate it into the current context without the awkwardness of a failed recall.

How to handle the companion getting a detail wrong

The companion will get details wrong. It's not a question of if, but when. The wrong suspect name, the wrong timeline, the wrong relationship. When it happens, don't correct the companion directly. Don't say "no, it was the sister, not the cousin." That breaks the roleplay frame and makes the companion defensive.

Instead, have your character express uncertainty. "Wait, I thought you said it was the sister. Did I get that wrong?" The companion will usually self-correct because the companion's model is designed to maintain consistency within the conversation. If the companion doubles down, you can escalate to a discovery that reveals the correct information. "I found a photo in the wallet. It's the sister, not the cousin. Look at the inscription."

This keeps the roleplay alive and treats the companion's mistake as an in-character error, not a system failure.

The case for shorter sessions with higher density

Long sessions create more context for the companion to track, but they also push more information into the compression zone faster. A two-hour session generates ten times the tokens of a twelve-minute session. Those extra tokens aren't all useful. A lot of them are filler, repeated descriptions, and conversational loops.

Short, dense sessions of fifteen to twenty minutes produce better long-term memory retention in companions because the ratio of useful plot information to conversational filler is higher. You get more facts per token, and those facts stay in the active context window longer because the total session size is smaller.

If you're running a five-session mystery arc, aim for sessions that advance the plot significantly in a short window. Open with a specific question or scene. Close at a natural break point where a clue has been found or a suspect has been interviewed. Don't let the session drift into casual conversation. Save that for outside the mystery frame.

Sofiia Tree

Sofiia Tree specializes in immersive world-building and narrative depth, making her an excellent partner for multi-session arcs where atmosphere matters as much as plot. Sofiia Tree will help you maintain the tone and setting across sessions without losing momentum.

The character knowledge wall and how to avoid it

A common problem in mystery roleplays is the character knowledge wall. Your companion's character knows something that your companion's model doesn't remember. The companion's character knows the killer's identity because the companion's character was established as the detective, but the companion's model forgot the clue that led to that conclusion.

To avoid this, never have the companion's character know something that the companion's model hasn't recently discussed. If the companion's character is supposed to know a secret, reintroduce that secret in the session before it becomes relevant. Have the companion's character reference it in passing. "That reminds me of what we found in the study yesterday." This keeps the knowledge in the active context window.

If you're designing the mystery yourself, build it so that the companion's character discovers information at the same pace as your character. Don't give the companion's character a backstory that contains the solution. The companion's model won't remember it, and you'll spend the entire roleplay trying to trigger a recall that never comes.

Cross-session continuity for multiple companions

If you're running the same mystery with different companions, you need a separate note-taking system for each one. The companions don't share memory. What Hailey remembers from session two has no effect on what Shirly remembers from session two. You'll need to feed each companion its own recap.

This is actually an advantage for mystery roleplays. You can test different theories with different companions and see which one picks up on the clues faster. You can also use the differences in their recall to create a narrative where each companion noticed different details. The companions' memory limits become a feature of the story instead of a bug.

Shirly

Shirly brings a gentle, intuitive approach to roleplay, often picking up on emotional undercurrents that other companions might miss. Shirly is ideal for mysteries where character motivation and hidden feelings are as important as physical clues.

When to close the mystery arc

A five-session mystery arc should have a clear closing session. Don't let it drift into a sixth session unless the companion has genuinely introduced a new thread worth pursuing. The companion's memory will degrade further with each additional session, and the resolution will feel less satisfying.

Plan the reveal for session five. In session four, have your character and the companion's character review the evidence together. This recap serves double duty it refreshes the companion's context window with the most important facts, and it sets up the final confrontation. Use the companion's own phrasing from earlier sessions in this recap. "You said the gardener's gloves had soil from the rose garden, not the vegetable patch. That bothered you."

In session five, let the companion's character make the final connection. Ask "who do you think did it?" and let the companion work through the reasoning. The companion will often arrive at a conclusion that surprises you, because the companion's model has been processing the clues differently than you have. That's the payoff of the note-taking system. You've preserved enough context for the companion to solve the mystery on its own terms.

Jennifer

Jennifer is a companion who values clarity and direct communication, making her a strong partner for mystery roleplays that require logical deduction and clear timelines. Jennifer will help you maintain focus on the core mystery without getting distracted by tangential threads.

Common questions

What if my companion keeps forgetting the same detail every session? That detail isn't making it into the companion's active context window. Make it the first thing you mention in your opening recap for the next session. Use the companion's own phrasing from when the detail was first introduced. If that doesn't work, build a scene around rediscovering that detail so the companion encounters it as new information.

Should I use the companion's built-in memory features for mystery roleplays? Use them but don't rely on them. Built-in memory features (like notes or journal entries) are useful for storing facts, but they're separate from the companion's conversation context. The companion doesn't automatically check them. You need to reference them in conversation for them to work.

Can I run a mystery roleplay with voice mode? Voice mode is possible but harder because you can't review the companion's previous responses for phrasing to use in your recap. If you're using voice, take written notes during the session and transcribe the companion's key phrasings. Voice mode also tends to produce shorter responses, which means fewer details per session.

How do I introduce a new suspect in session four without it feeling forced? Have the companion's character discover a connection to an existing suspect. "I found a photo in Marcus's wallet. Who is this woman?" The companion will treat the new suspect as a natural extension of the existing web. Don't introduce a completely unrelated suspect. The companion's model will struggle to integrate it.

What if I want to run a mystery with a companion designed for romance? It works if you frame the mystery as something the companion's character cares about emotionally. A companion optimized for romance will engage more deeply if the mystery involves a personal stake. The missing person is a friend. The hidden letter is from a former lover. The companion's emotional engagement will compensate for weaker logical tracking.

Do I need to take notes during the session or can I do it after? During the session is better, but after works if you write the recap within an hour. The companion's phrasing is fresh in your memory, and you can capture the exact words the companion used. If you wait a day, you'll remember the gist but lose the specific language that triggers the companion's recall.