Try AI Dungeon free before buying a longer-context plan. It suits readers who want to take turns inside a changing story. Our logged-out web session generated a short continuation, but its model selection, memory switches and exact-detail handling all needed inspection. A fluent opening is a starting point, not evidence that a long adventure will stay consistent.
One guest continuation worked; the inscription needed checking
On October 26, 2026, Play → Quick Play opened a custom guest adventure. The page offered sign-up to save the game. We did not create an account or test saved-game recovery.
We supplied a calm scene about Mara, a fictional 35-year-old archivist, finding a brass key engraved WEST and checking an archive door. We asked the AI to introduce no other characters.
Actual check · Logged-out web guest · One generation
The continuation stayed with the key and archive door, without introducing another character. It referred to an engraved “W”, while our input specified WEST.
I would flag that wording for review rather than assume the full inscription was preserved. It could be a shortened reference, not necessarily a contradiction. One response does not establish a failure rate or long-term recall.
That is the useful way to read a trial: check one or two facts that matter to your story, alongside whether the prose is enjoyable. We did not measure speed, repeat the sample or compare paid outputs.
Dynamic Small is a rotation, not one fixed model
Our model selector displayed Dynamic Small, version 1.0.0. The official model guide describes Dynamic Small as a system that randomly switches among models; the underlying mix is not published as a fixed list.
A different retry is not automatically evidence that your prompt improved. The model behind the response may also differ. For a controlled comparison, use a named model if your access permits it, and keep the story, settings and context constant. We could not identify which underlying model produced our guest response.
Inspect memory switches before trusting the Memory System label
In our web session, the path was Game settings → Gameplay → Models → Memory System. Before generation, it showed:
- Context length
- 4,000 tokens
- Auto Summarization
- Off
- Memory Bank
- Off
These are the settings we observed, not a promise about every account or future default. We left them unchanged. Our short continuation therefore did not test the automatic memory features.
The Memory System guide separates a running summary from retrieval of relevant stored memories. A stored memory is not necessarily included in every response; retrieval and the available context budget matter.
Keep important facts compact. Context also holds instructions, plot material and recent story text. The context-construction guide explains that overfilling the allocation for required material can cause trimming or exclusion. Treat “always in context” shorthand as a priority rule, not unlimited storage.
When a detail goes missing, inspect the latest context before adding another long paragraph of instructions. The settings guide documents Inspect Input and View Context for checking what was sent for a recent output. We did not test those inspection tools in this session.
Journey includes different context for different models
The membership page listed Journey at USD $14.99 per month when checked October 26. That is its published monthly price, not an annual equivalent or a checkout price we tested. The page says subscriptions renew automatically.
The model guide lists this comparison for Journey:
Source: current model-specific allowances. The same subscription does not mean the same context on every model. Neither number is a quality score, and we did not benchmark either paid configuration.
For Dynamic Large, the guide also lists optional extension at +4K per credit per action, up to 32K total. At that setting, extending 30 actions by one increment would use 30 credits for the extension alone. That is arithmetic from the published rule, not a measured bill or a dollar estimate.
Before paying, choose the model you expect to use and check its included context and extension charge. Do not buy from a tier’s largest advertised context number alone.
Search ratings do not set the AI’s generation safety
In Discover → Filters, our logged-out view had Everyone and Teen checked and Mature unchecked. Inside the guest adventure, Gameplay → Models → Safety separately showed Moderate selected. We did not change either control.
The settings documentation says a scenario’s rating does not set AI Safety. It also describes Safe as the default, which differed from the Moderate selection visible in our guest session. Read the active control rather than assuming a documented default. This observation does not establish a universal default change.
Browsing labels classify community content; AI Safety governs generated text. Neither is a privacy setting. Use fictional material for the trial; Latitude’s privacy policy separately describes service-provider sharing and public user content. We did not audit data handling or publish our test adventure.
Next: make the free trial produce a useful test record
Before an upgrade, try another short fictional scene and record:
- The selected model or rotation label, plus the displayed version.
- Context length and the state of the two memory controls.
- Two exact story facts to check in the continuation.
- Any correction you needed, and whether repeating the check actually helped.
Choose AI Dungeon if taking actions inside a story feels useful after that exercise. If your main task is editing manuscript continuations, compare the existing NovelAI review; for steering a cast and scene setup, read the DreamGen review. Pick the workflow first, then assess the paid model and context it requires.