Why AI roleplay forgets at all
A language model doesn’t remember your story; it reads it. Before every reply, the app gathers text (its instructions, the world, the recent story, and whatever else it chooses) and sends it along with your newest turn. The model writes from that and nothing else, and when the reply is done, nothing stays behind. The next reply starts from a fresh read.
So “the AI forgot” always means one of two things. Either the fact wasn’t in the text sent with that reply, or it was there and the model didn’t use it well. Every fix in this guide targets one of those two failures.
The first failure is far more common, and it has a hard cause: the context window, the most text a model can read for a single reply. A long roleplay outgrows any window. Once it does, the oldest scenes stop being sent, and as far as the model can tell, they never happened.
Six kinds of forgetting, and what causes each
Forgetting shows up in a few recognizable ways, and each points to a different cause. Match the symptom before you reach for a fix:
| Symptom | Usual cause | First fix |
|---|---|---|
| A name changes or vanishes | The scene that introduced it has left the window, and nothing else carries the name | Give the person a lore card with their name and nicknames as tags |
| A wound heals or a locked door opens itself | Small physical facts lived only in old turns, and summaries tend to drop them | Keep facts that matter right now in session notes |
| A personality drifts | Recent replies now outweigh the definition, and the model imitates its own latest lines | Regenerate drifted replies before you answer them |
| A promise or plot thread disappears | The memory exists, but nothing in the current scene calls it up | Mention it by name, or check the recap |
| A contradiction two messages later | Not memory at all: ambiguity, crowding, or instructions that disagree | Regenerate, and remove whatever conflicts |
| A new chat remembers nothing | A new session starts with no memories | Continue in the original session |
The rest of this guide is the machinery behind that table: what the window holds, what a memory system adds, and how to keep each layer healthy.
The context window is the hard edge
Windows are measured in tokens, the small pieces of text a model reads, at roughly three quarters of an English word each. Everything sent with a reply shares the window: instructions, the world’s settings, your character, lore, memories, the recent story, and room for the reply itself.
On LoreWeaver, your plan sets the size of the window and how many memories one story can keep:
| Plan | Context allowance | Roughly, in words | Memories per story |
|---|---|---|---|
| Free | 6,400 tokens | 4,800 | 30 |
| Plus | 16,000 tokens | 12,000 | 300 |
| Pro | 32,000 tokens | 24,000 | 1,200 |
| Ultra | 64,000 tokens | 48,000 | Unlimited |
| Legendary | 100,000 tokens | 75,000 | Unlimited |
Even the largest window fills. A campaign that runs for months can grow longer than a novel, far more than any window holds word for word. When the recent story no longer fits, the oldest turns leave first, and from then on older events can reach the storyteller only through the other layers.
Why bigger isn’t the whole answer
A bigger window keeps more of the recent story exact, which helps most in dense scenes. But long prompts take longer to answer, and models tend to use the beginning and end of a long prompt more reliably than its middle (Lost in the Middle). What goes into the window matters as much as how much fits.
What a memory system adds
A memory system works around the window instead of trying to enlarge it. It does three jobs:
- It writes things down. As the story goes, it saves short notes on what happened: who was there, what changed, what was promised.
- It finds the right notes again. Before each reply, it searches those notes for the ones that fit the moment and sends only those.
- It keeps the arc in view. A running summary covers the whole story in a few paragraphs, so big events stay present even when the scene isn’t about them.
Lore does a fourth job that is easy to confuse with memory. Memory records what happened; lore records what is true, such as who a person is, what a city looks like, and what a faction wants. A long story needs both, because a promise and the person it was made to are different kinds of fact.
Each job can fail in its own way. A note can be wrong, a search can miss, and a summary can flatten a detail that turned out to matter, which is why the fixes below include reading and correcting what the system wrote.
How LoreWeaver keeps a long story together
LoreWeaver does all four jobs as you play, and lets you see and correct each one:
- Recent turns, word for word, as many as your plan’s window allows.
- Memories, written every few replies from the story itself. Each records the people involved, keywords, a category, and an importance score, and the writer is shown the names already established, so “Marcus” doesn’t drift into “Mark.”
- Recall that weighs meaning first. Memories are found by what they mean, which carries about half the score, then by shared wording, by names and keywords, and with a small lift for importance. Age plays no part, so a vow from the first week competes on equal terms with yesterday’s argument.
- The recap, which reads each new stretch of the story and folds it in: the story so far, and where things stand now, with every open promise and deadline. The storyteller reads it on every reply.
- Lore cards, which join a reply when they’re pinned or when their title or a tag appears in the recent conversation.
- Edits that carry through. Rewind or undo a scene, and its memories, recap changes, and Loom changes go with it; branch a story, and the branch starts from what was true at the fork.
Each plan sets how many memories one story can hold, from 30 on the free plan to no limit on Ultra and Legendary. When a story reaches its plan’s limit, it stops adding memories, and the story is carried by its recap, its lore cards, and the turns still in view. Memories you write yourself don’t count toward the limit.
Fixes, from quickest to deepest
Most forgetting takes a minute to fix once you know which layer missed. In rough order of effort:
- Say the name. “I ask Anya about Tomas” finds memories and lore that “I ask her about her brother” never will.
- Regenerate a contradiction before you answer it. Once you play on from a reply that gets a fact wrong, it becomes part of the story, and later memories are written from it.
- Pin down what matters right now. Injuries, disguises, who knows which secret: a line in your session notes keeps them in front of the storyteller until the scene is over.
- Give recurring people and places a lore card. One card with a name and its nicknames as tags brings them back on cue for the rest of the story.
- Read the recap after turning points. Edit whatever it got wrong, and fix the memory underneath too, because the recap is rebuilt from memories and a rebuild can overwrite your edit.
- Let big moments happen in the scene. Memories are written from the story itself, so a vow spoken in a scene is remembered better than one you only meant.
- Check the context inspector on the reply that forgot, before you change anything.
- Keep one campaign in one session. Each session has its own memory bank, so branch for what-ifs instead of starting over.
- Give a long campaign more room. A larger plan keeps more of the recent story word for word, and more memories per story.
Four myths about AI memory
- “The AI learns my story as I play.” It doesn’t. Everything a storyteller seems to remember is text the app sends back with the next reply, which is exactly why you can read it, edit it, and rewind it.
- “Memories are a transcript.” They are short notes on what mattered. Small physical details often don’t survive the summary, which is why lore cards and session notes hold the specifics.
- “Repeating important facts in every message helps.” It crowds the recent story with the same lines, and the storyteller may start repeating them back. State a standing fact once, where the storyteller reads it when it matters: your persona or session notes for every reply, a lore card for whenever its name comes up.
- “A new chat clears up the confusion.” It clears everything. A new LoreWeaver session starts with no memories at all, though the world’s lore cards carry over, so start fresh only when you want a fresh story.
