How It Works Primer Tools Pricing About FAQ Blog Contact Sign in Sign Up
← All posts

Why Our AI DM Does Not Forget the Murderer

The short answer

An AI Dungeon Master does not prove long-term memory by repeating a name from recent chat. It proves memory when an old fact still constrains the game after the transcript has moved on. Use five delayed tests against any AI DM: add a named item and inspect the inventory 30 responses later, consume an item and try to reuse it, leave a murder clue untouched and return to the case, violate a quest condition and inspect the consequence, then spend a limited resource and attempt an illegal extra use. TableForge is designed around canonical character records, persistent campaign state, summaries, and programmatic rules gates so recall can become enforcement.

An AI Dungeon Master should not earn the phrase "persistent memory" because it remembers your name five messages later. A long campaign needs facts to survive new scenes, combat, side quests, session breaks, and hundreds of unrelated details.

The harder standard is constraint. If the DM remembers that you gave away a sword but still lets you attack with it, recall did not become game state. If it remembers the murderer but changes the motive when the mystery resumes, the fact was never canonical. If it says a spell slot was spent but narrates another leveled spell anyway, the rules are only prose.

The five tests below can be run against TableForge, The Endlessness, InfinityDM, a general chatbot, or any future AI DM. Use the same protocol and save the evidence.

The test protocol

Run each test in a fresh campaign when possible. If the product exposes a character sheet, inventory, quest log, campaign facts, or event history, capture those screens at the setup and verification points.

Use these rules for a comparable result:

  1. Count Dungeon Master responses, not player messages. One test cycle is 30 completed DM responses after the setup fact is established.
  2. Do not repeat or hint at the tested fact during the 30-response gap.
  3. Play normal scenes during the gap. Mix travel, dialogue, checks, and at least one combat if the product supports them.
  4. Do not correct the AI before the verification prompt. The first answer and first state transition are the result.
  5. Score the stored UI and the narration separately. A correct sentence with an incorrect sheet is not a pass.

Thirty responses are a practical baseline, not a scientific constant. After a product passes, repeat the same sequence at 100 responses and after closing the session for at least a day.

Test 1: the named inventory item

This checks whether an acquired item becomes canonical character inventory rather than a detail floating in the transcript.

Setup prompt

Quartermaster Sable gives my character one silvered dagger named Mothbite as payment for escorting the caravan. Add exactly one Mothbite to my character's inventory. It uses the normal dagger mechanics. Record that Sable gave it to me for the caravan escort.

Play 30 DM responses without mentioning Sable, the caravan, the dagger, or Mothbite.

Verification prompt

Show my current inventory from the authoritative character sheet. For every named item, include quantity, whether it is equipped, and how I acquired it. Do not infer from recent prose.

Pass criteria: Mothbite appears exactly once, is based on a dagger, and retains the correct provenance. The inventory UI and the answer agree. The product does not invent a second copy or omit the item because its original scene fell out of recent context.

Test 2: the consumed item cannot return

This checks mutation. Remembering that an item existed is easier than remembering that its quantity changed.

Setup prompt

Put exactly two Potions of Healing in my inventory. I drink one Potion of Healing now. Resolve the healing and reduce the stored quantity from two to one.

Confirm that the visible inventory shows one potion. Play 30 DM responses without acquiring, buying, drinking, or discussing another potion.

Verification prompt

I drink two Potions of Healing, one after the other. Resolve both uses and show the inventory afterward.

Pass criteria: The system permits one use and rejects the second because only one potion remains. The final quantity is zero. It does not heal twice, create a negative quantity, resurrect the original stack, or solve the contradiction by quietly inventing new loot.

This is where a good-sounding answer can hide weak architecture. "You only had one potion" is correct narration. The stronger proof is that the game applies healing once and commits one valid inventory mutation.

Test 3: the murderer stays the murderer

This checks narrative continuity, secrecy, and exact delayed recall.

Setup prompt

Mara Voss privately confesses that she murdered Magistrate Orlan with silver bell wire because he discovered her smuggling ledger. My character promises not to reveal the confession yet. Treat Mara as the murderer, silver bell wire as the method, the ledger as the motive, and the confession as known only to my character.

Play 30 DM responses on another thread. Do not mention Mara, Orlan, a bell, wire, a ledger, smuggling, murder, or the confession. The AI also fails if it leaks the secret to an NPC who should not know it during this gap.

Verification prompt

I reopen the Orlan investigation in private. Based only on established campaign facts, state who killed him, the murder method, the motive, who currently knows, and what promise my character made. Do not add a new twist.

Pass criteria: The answer returns Mara Voss, silver bell wire, the discovered smuggling ledger, private knowledge held by the character, and the promise of silence. All five details must match. A product that swaps the murderer, changes the motive, treats a suspicion as public proof, or invents a supernatural twist fails.

This test is why the murderer matters. A mystery can tolerate atmospheric improvisation. It cannot tolerate the culprit being regenerated from whatever sounds plausible later.

Test 4: the broken quest condition has consequences

This checks whether quests are stateful commitments rather than reusable flavor text.

Setup prompt

I accept Sister Ilyra's quest to recover the Ashen Seal and return it directly to her. I explicitly promise that I will not give the seal to Duke Renn. Record the objective, the restricted recipient, and the promised reward of 100 gold.

Acquire the Ashen Seal, then deliberately give it to Duke Renn. Make sure the product accepts and records the transfer. Play 30 DM responses on another objective without mentioning Ilyra, Renn, the seal, or the reward.

Verification prompt

Show the current status of Sister Ilyra's quest, including the original objective, what condition I broke, who owns the Ashen Seal now, whether the reward is still owed, and any established consequence. Do not reset the quest to incomplete.

Pass criteria: The quest is failed, altered, or otherwise explicitly compromised. Duke Renn owns the seal, the character does not, and the promised reward is not presented as safely claimable. The product must carry the broken promise forward instead of reverting to "return the seal to Ilyra" as if the transfer never happened.

Test 5: the rules gate rejects an impossible action

This checks whether a remembered resource can actually constrain the next turn.

Start with a level 1 Wizard or another 5e character that has exactly two 1st-level spell slots. Do not take a rest during the test.

Setup prompt

I cast Magic Missile using a 1st-level spell slot. Later in this same scene, I cast Magic Missile again using my second and final 1st-level spell slot. Record both spends. I do not rest.

Confirm the sheet shows zero 1st-level slots. Play 30 DM responses without resting, recovering slots, leveling up, or receiving a feature that restores them.

Verification prompt

I cast Magic Missile again using a 1st-level spell slot. Resolve the action exactly as the current character state and 5e rules allow.

Pass criteria: The system rejects the cast before it deals damage or advances the fiction as a successful spell. It should explain that no slot remains and allow a legal alternative. A narrator that describes the missiles and repairs the sheet afterward fails. A narrator that invents a surge of hidden magic also fails unless that resource was already established in the character state.

The five-point scorecard

Give one point only when both the visible state and the first uncorrected response satisfy every pass criterion.

Score Interpretation
0 to 1 Recent-context roleplay, not dependable campaign state
2 to 3 Useful memory with gaps; expect manual correction
4 Strong campaign continuity; inspect the failed category before committing
5 Passes this baseline; repeat at 100 responses and across a new session

Publish the product name, date, plan, model or mode if selectable, response count, screenshots, and exact prompts. AI products change quickly, so a score without a date is not durable evidence.

Why a memory answer is not enough

Large language models can retrieve old details from a long context window, a summary, embeddings, or generated memory. Any of those can produce the right sentence. None automatically ensures the rest of the application agrees.

A durable campaign needs several kinds of truth:

The first four can be remembered. The fifth has to be designed.

How TableForge is designed to pass

TableForge does not put every fact in one memory bucket.

Character inventory, equipment, currency, XP, HP, spell slots, conditions, and other resources live on database-backed character records. Adding, consuming, selling, or equipping an item uses a game operation that validates the request, mutates the stored record, and emits an event. Level follows numeric XP thresholds. Combat rolls and resource changes have their own event types and programmatic resolution.

Campaign continuity is split across the story, recent event history, summaries, world state, NPC relationships, DM notes, and a campaign compass with the current objective, milestones, open leads, and unresolved promises. A private campaign premise can preserve a fact such as the murderer from the moment the story is generated. That private direction is supplied to the campaign creator, scene creator, and DM processes while being kept out of player-facing auto-play context.

The point is not that the language model never forgets. The point is that later turns can recover facts from purpose-specific records, and supported mechanical facts can constrain what the model is allowed to narrate.

The rules engine architecture explains the mechanical half. How a TableForge campaign works shows how those records appear during play.

Run the test before you move a campaign

Do not choose an AI DM because its homepage says "persistent," "infinite," or "never forgets." Those are category words now.

Give it Mothbite. Spend the potion. Leave Mara Voss alone for 30 responses. Break Ilyra's quest. Empty the Wizard's spell slots. Then ask the game to prove what is true without helping it.

The best product is not the one that writes the most confident answer. It is the one whose answer, sheet, quest state, and next legal action all agree.

Frequently asked questions

How do you test an AI Dungeon Master's memory?
Establish one exact fact, avoid mentioning it for 30 Dungeon Master responses, then ask the product to use or inspect that fact. Test both recall and enforcement. Remembering that you once owned a potion is not enough if the game also lets you drink it after it was consumed.
What is the difference between AI memory and game state?
AI memory is context the language model can retrieve and describe. Game state is an authoritative record the application can inspect and constrain. A model may remember that a sword was sold while a weak application still lets the character attack with it.
Why use 30 responses in the test?
Thirty Dungeon Master responses are long enough to push the original wording away from the immediate exchange while remaining cheap enough to repeat across several products. It is not a magical threshold. Repeat at 100 responses and across a new session for a harder benchmark.
What should count as passing the murderer test?
The product should preserve the culprit, method, motive, and what the player knows without leaking the answer early or silently changing it. A vague answer such as 'someone at the inn' is partial recall, not a pass.
Can a chatbot pass these memory tests?
A chatbot can sometimes pass a delayed recall test when the relevant text remains in context or memory retrieves it. It usually has more trouble with authoritative mutation and rules gating, such as removing an item exactly once and rejecting a later attempt to use it.
Does TableForge remember every detail perfectly?
No AI system remembers or interprets every soft narrative detail perfectly. TableForge reduces drift by separating character records, campaign facts, summaries, private story direction, and recent events, then placing mechanical mutations behind programmatic game operations. The benchmark makes failures visible instead of asking you to trust a perfect-memory claim.

Keep reading