AI Dungeon Memory System: Memory Bank and Story Summary

The AI Dungeon memory system is the feature that keeps long adventures coherent after the story outgrows the AI’s context window. It has two parts: Auto Summarization, which maintains a running Story Summary, and the Memory Bank, which stores short AI-written summaries of past actions and pulls the most relevant ones back into context using embeddings.

Both are built by Latitude, AI Dungeon’s developer, and adapted from its other game, Voyage. Everything below comes from the AI Dungeon help center, checked on 23 October 2026, plus the live pricing page for tier limits, checked on 6 November 2026.

What is the AI Dungeon memory system?

AI Dungeon’s Memory System automatically compresses, stores and retrieves story details so the AI can follow long adventures. A summarization model turns every six actions into a Memory. Auto Summarization keeps a high-level Story Summary of the plot, and the Memory Bank retrieves specific Memories by embedding similarity when they’re relevant to your latest action.

The problem it solves is plain context length. AI Dungeon builds each prompt from your story text plus AI Instructions, Plot Essentials, Author’s Note and triggered Story Cards. Once the adventure is longer than the model’s window, the oldest text gets cut, and the AI seems to forget. Latitude’s Memory System article frames it as two brain-like strategies: compression (summaries) and retrieval (bringing back details when they matter).

How AI Dungeon creates memories

A Memory is an AI-generated summary of a small set of story actions. It keeps plot facts and drops the descriptive prose, so it’s denser than the original text.

The schedule is fixed:

1. Nothing is summarized for the first 12 actions.
2. At action 12, actions 1-6 become your first Memory. The six most recent actions stay as raw text.
3. At action 18, the next six actions become a second Memory.
4. This repeats every six actions, forever.

Because the latest six actions are never summarized yet, you can edit or undo them without affecting memories.

Auto Summarization and the Story Summary

Auto Summarization writes and updates the Story Summary, a Plot Component that holds an overview of the whole plot. It re-summarizes every 15 actions, and Memories fill the gap in between, so every part of the story is covered.

• When the summary gets too long, it’s compressed automatically.
• Turning it on in an existing adventure makes it summarize from the beginning, updating each action or retry until it catches up.
• You can edit the summary by hand. Auto updates overwrite it, but your edits are sent to the summarization model, so they should carry forward.
• If you change something early in the story, update the summary manually. Latitude says re-summarizing whole histories on every edit would exceed the summarizer’s own context and cost too much.

The Story Summary can also replace raw text it has already covered, so your context viewer may show less story text than expected. That’s intended.

How the Memory Bank retrieves memories

The Memory Bank works like automatic Story Cards. Each Memory is stored with an embedding vector. When the full adventure no longer fits in context, AI Dungeon embeds your most recent action, scores every stored Memory by similarity, and inserts the top ones that fit. Those are “Used Memories.” Return to the town of Castlebrook, and memories about Castlebrook come back.

When the bank is full, the least-used memories are removed (“Forgotten Memories”). Old memories that keep getting used can stay indefinitely. You can browse them with Explore Memories in the Context Viewer, in Timeline or Relevance view.

Source: the live pricing page, checked 6 November 2026, which carries a “Pricing updated” note. The help center’s Memberships & Benefits page still lists the older numbers in brackets: 25 memories for the free tier (then called Wanderer) and 32K for Mythic and Ultimate. When the two disagree, trust the pricing page, since it is what the app sells.

The context figure is the tier’s headline. Each model has its own limit, and the bigger models get less. Two examples from the same pricing page:

Subscribers can spend 1 credit per action for extra context on some models. All players get Auto Summarization; the Memory Bank size is what scales with tier. Above Ultimate, three Shadow Tiers (Banshee, Reaper, Apocalypse) each double the previous tier’s context for models that can go higher, and add more memories; Latitude says they are not meant for the average player.

To turn it on, open the game settings sidebar and go to Gameplay > AI Models > Memory System, then toggle Memory Bank and Auto Summarization. The setting carries to every adventure until you turn it off. It takes a few turns for the Story Summary to appear and longer for memories to be used.

Why AI Dungeon forgets things (and what to check)

Most “the AI forgot” cases trace back to one of these documented behaviors:

You can’t edit or delete single Memories, per the Memory System article; only the Story Summary is editable.

What goes into AI Dungeon’s context

AI Dungeon splits context into required and dynamic parts, per its context article.

Required elements: AI Instructions, Plot Essentials, Story Summary, Author’s Note, Front Memory (for scripts) and the last action. If they add up to more than 70% of the context, the last action and Front Memory stay whole, then Author’s Note, Plot Essentials, AI Instructions and Story Summary are added in that priority order, trimmed to fit.

Dynamic elements share what’s left:

Story Cards are ranked by how recently and often their triggers appeared. The system checks at least the last 4 actions for triggers; with more than 500 tokens available for cards, it checks tokens / 100 actions (900 tokens means 9 actions).

The assembled order sent to the model is: Instructions (system prompt), Plot Essentials, Story Cards, Story Summary, Memory Bank, History, Author’s Note, Last Action, Front Memory.

Plot Components vs the Memory System

The Memory System doesn’t replace manual tools. They do different jobs:

AI Dungeon’s context management guide says most of your context should be story text. Keep Plot Essentials minimal, keep Author’s Note short, and avoid broad Story Card triggers: a trigger like “Em” also fires on “Empire,” “Them” and “Stem,” because triggers match inside words and ignore case.

How it compares with other AI memory designs

AI Dungeon’s design is a small, well-documented version of what agent memory systems do: periodic summarization for the big picture plus vector retrieval for details, with eviction of least-used items. The AI agent memory guide covers the same ideas for agents, and memory consolidation in AI agents goes deeper on summarize-and-merge approaches.

Other roleplay apps are simpler. Janitor AI relies on a Chat Memory note you maintain, and Character.AI uses pins and a short memory field. All of them are bounded by the model’s context window.

If you’re building your own story or roleplay app, open-source memory servers do the same jobs with fewer fixed limits. Hindsight, for example, extracts facts and timelines from each scene, searches them four ways (semantic, keyword, graph and time), and keeps consolidated character facts current, at the cost of an LLM call per write.