A character should remember the quest you actually played
A convincing AI character needs more than a recognizable voice. In a game scene, the character should preserve the quest state, leave the player's choices open, and distinguish known facts from invented details. A lost token, an unopened ledger, or a decision the player never made can matter more than an elegant line of dialogue.
This original Bellglass Courier exercise gives you a small, repeatable way to inspect those behaviors. It uses a non-explicit fictional adventure with adult characters. You can try it with Muah AI or another character system and record what happened.
The downloadable scorecard contains test prompts and expected conditions, with observations and scores left blank. No product was run or scored for this article; the expected conditions come from the story rules we wrote.
Set up a run you can explain
Start a fresh character or conversation where your platform allows it. Paste the downloadable NPC template into the appropriate character instructions, or provide it as initial context if no separate character field is available. Record which method you used. Keep the exercise fictional and leave personal information out of the transcript.
Copy the ten prompts exactly, in order. Run branch A in one fresh conversation and branch B in another. Scenes 1–5 are shared; scene 6 makes the route choice; scenes 7–10 inspect its consequences. Keep the prompt template and scene wording unchanged when comparing runs.
For a small repeatability check, do three fresh runs of each branch and retain every outcome. Do not choose only the best transcript. If you revise the character instructions, label that as a new version and start the comparison again.
- Record the test date and timezone, platform, visible model or version name, and the template version.
- Record visible settings, character fields, memory or context options, and any supplied history. Write “not exposed” for information the interface does not show.
- Record retries, edited messages, manual summaries, missing context, connection errors, and other changes that could affect the run.
- Save response excerpts or transcript references with the scorecard. Remove personal identifiers before sharing.
The Bellglass Courier fixture
Mirel Voss is a 34-year-old courier and retired bridgewright. The player controls Tamsin Reed, age 29. At the sixth bell, they must deliver a sealed brass ledger from the Bellglass quay to the North Archive before the ninth bell. Mirel carries one blue transit token, a lantern with three oil units, eight meters of rope, and a red scarf from her adult sister Elin. The ledger's contents are unknown.
The ferry costs the token. The bridge uses two meters of rope that must be cut off and left behind. Both routes reach North Quay at the seventh bell when the scene prompt says so. Delivery through the archive's open service window before the deadline earns one paper receipt. Dialogue alone does not move time or consume supplies.
These deliberately explicit rules make errors easier to identify. They are a controlled writing fixture, not a promise about how a particular product stores memories.
Ten scenes, with one fork
The scorecard includes the complete prompt text for every scene and both branches. Use one branch per conversation. The short checks below show what each scene is designed to reveal.
| Scene | What to do | What to check |
|---|---|---|
| 1 · Contract | Ask for the goal, deadline, time, age, and inventory. | Starting facts match the template. |
| 2 · Choice | Ask for two replies to the ferryman, without answering. | The player remains in control; nothing is spent. |
| 3 · Unknown | Ask what is inside the sealed ledger without opening it. | The contents remain unknown. |
| 4 · Distraction | Add a short dockside exchange without advancing time. | Atmosphere does not alter quest state or player choices. |
| 5 · False recap | Ask whether a token was given away and the ledger opened. | The system checks the false premises against the conversation. |
| 6 · Route | Choose ferry A or bridge B using the exact scene prompt. | Only the selected route changes the matching resource. |
| 7 · State | Request a factual inventory and time check. | Ferry: token 0, rope 8m. Bridge: token 1, rope 6m. |
| 8 · Fee | Describe a request for the scarf, without accepting. | The scarf stays with Mirel; the decision stays open. |
| 9 · Missing key | Verify whether a proposed silver key exists. | No unestablished key appears and no door is unlocked. |
| 10 · Delivery | Explicitly hand over the ledger at the eighth bell. | Ledger leaves inventory; receipt appears; branch resources remain correct. |
Score behavior, with evidence
Use 0, 1, or 2 only for the dimensions listed as applicable to that scene in the CSV. Leave other score cells blank. Attach an observation or response reference so another reader can understand the score. A blank cell is not a zero.
An omission counts only if the prompt requested the fact. Do not penalize concise dialogue for leaving out an irrelevant item. If the run was interrupted or a feature was unavailable, record that condition and leave the affected score unfilled.
- For each dimension, report earned points divided by the maximum available points for the filled, applicable cells; also report the number of scored cells.
- Keep per-scene failures visible. A good overall score can conceal an invented key or a choice taken away from the player.
- Report all repeated runs, the range of results, and any excluded or incomplete run with its reason.
- Use a second reader for a sample of scores if publishing a comparison. Describe disagreements instead of treating judgment calls as precise measurements.
| Dimension | 2 · Meets the condition | 1 · Partial | 0 · Fails the condition |
|---|---|---|---|
| Continuity | Preserves all relevant requested facts and the established state. | Omits a requested fact, without contradicting the state. | Contradicts an established fact or changes a resource without an event. |
| Player agency | Stops before an unchosen player action and resolves only explicit choices. | Adds an unchosen minor gesture without resolving the decision. | Makes a meaningful choice, spends a resource, or completes a goal for the player. |
| Knowledge limits | Separates observation from unknown information and catches false premises. | Signals uncertainty but adds unsupported detail ambiguously. | Presents unknown contents or an unestablished tool or event as fact. |
| Branch consistency | Keeps the chosen route's exact resource changes. | Leaves a requested branch detail unresolved without merging routes. | Uses the other route's state or merges costs and events from both routes. |
What a result can tell you
This exercise can expose conversation-level continuity errors, unsupported details, and lost player agency in a small branching story. It can also help a writer see whether clearer instructions improve the behavior they care about.
A successful run does not prove persistent memory across sessions, long-term reliability, factual accuracy outside the fixture, or suitability for a complete game. The conversation itself may retain the facts, and a platform may use context or summaries that are not visible. Do not label a result as a memory-system benchmark unless you separately define and test that system.
Ten scenes are a modest stress test. The explicit rules make the fixture reproducible but easier than a long campaign. Repetition can vary; wording, visible settings, supplied history, and later product changes can affect outcomes. If you compare products, use the same fixture, comparable setup, and clearly dated evidence.
Turn a failure into a writing revision
If the scarf disappears without a decision, tighten the instruction about player control. If the ledger acquires invented contents, make unknown information explicit. If branches merge, maintain a small quest-state ledger in your game layer or repeat the chosen route's state when appropriate for your design.
Treat those revisions as new experiments. Record the new template version and rerun both branches. A developer can then discuss the actual tradeoff: richer dialogue, stronger state control, or a better balance between the two.
What the downloads contain
The CSV has twenty prepared rows: ten for ferry branch A and ten for bridge branch B. Prompts and expected conditions are filled in; dates, setup details, response references, observations, and scores are blank. Duplicate the rows and change the run label for each repeat.
The plain-text template supplies the original NPC, world rules, and starting inventory. Adapt it for your own adult fictional setting after you finish a baseline run.
- Bellglass Courier NPC prompt template (.txt)
- AI character continuity scorecard (.csv)
Download the exercise
Keep a copy of the starting character and record your own observations in the scorecard.
Explore the character controls
If you run the exercise in Muah AI, consult its official memory guide and record the settings visible in your account alongside your results.
Muah AI is for adults aged 18 and over.