Every chat model reads a story through a window: the recent conversation is vivid, the distant past fades, and beyond the window's edge, nothing exists. Long, satisfying stories in Janitor AI on Android are built by users who understand that window and manage it deliberately — refreshing key facts, keeping summaries, and treating forgetfulness as weather rather than betrayal.
This guide explains how context behaves in practice, the summary habit that carries a story across hundreds of messages, which facts deserve active maintenance, and how to recover a scene when the model contradicts something established.
None of this requires technical knowledge. It requires the same skill a good serial novelist uses: reminding the reader — here, the model — of exactly what matters, exactly when it matters again.
How the Window Actually Feels
Recent messages dominate: the last exchange shapes tone, facts, and momentum. Older material survives in proportion to how often it is re-mentioned, which is why a name used every ten messages persists and a crucial detail mentioned once, two hundred messages ago, evaporates.
This is not a flaw to rage against but a physics to design around. Human conversation works similarly — we repeat what matters — and the techniques below simply make that repetition graceful instead of clumsy.
The Summary Habit
Every so often — at a chapter break, a new location, a time skip — pause and write or request a short summary of the story so far: who, where, what was decided, what is unresolved. That summary, sitting inside the window, re-anchors everything it names.
Keep your own running summary in a note as well, and paste or adapt it when starting fresh sessions of the same story. Users who do this report stories that survive for months; users who do not are re-explaining their world weekly. The settings that support long sessions are covered in the Janitor AI advanced settings guide, and summaries are the human half of that machinery.
Facts Worth Actively Maintaining
Maintain the load-bearing facts: names and relationships, promises and debts, injuries and possessions, the rules of the world, and any secret one character holds from another. Decorative facts can fade harmlessly; load-bearing ones collapse scenes when they go.
Refresh a fact just before it matters again — the chapter where the debt comes due gets a quiet reminder in an early message. Timing a reminder is the difference between stagecraft and nagging.
Scene Craft That Helps the Model
Enter scenes with the essentials restated naturally: place, who is present, the emotional weather carried over from last time. A single grounding sentence at a scene's head re-loads more context than any setting, because the model builds on what the conversation itself asserts.
Close scenes deliberately too. A short beat acknowledging what changed — a decision taken, a door closed — gives the next scene a foundation the model can actually find when it looks back.
When the Model Forgets or Contradicts
Correct briefly and in-world where possible: a line in character reminding of the fact teaches the model and preserves the scene simultaneously. Save out-of-character corrections for the stubborn cases, and keep them short; the story is the patient, not the argument.
If a contradiction reveals the model has lost a whole thread, do not wrestle — re-anchor. Paste your summary, restate the immediate situation, and continue from solid ground. Recovery by re-anchoring succeeds nearly always; recovery by arguing succeeds mostly in ruining the evening. The beginner-facing version of these habits lives in roleplay tips for beginners in Janitor AI, which introduces them gently.
One Long Chat or Many Linked Ones
A single endless chat accumulates drift; linked chapters with fresh summaries accumulate story. Prefer the chapter model for anything you care about: end sessions at natural breaks, carry the summary forward, and let each chapter start with a window full of what matters.
Chaptering also protects you from the rare catastrophic loss. A story distributed across saved summaries can survive an app problem that would take a single monolithic chat with it.
Memory and Model Choice
Different models handle long context with different grace, and larger context windows genuinely help epics. If your stories are long, weigh context behaviour when choosing your configuration — the settings guide in this library explains where that choice lives.
But technique beats hardware here as everywhere: a well-summarised mid-sized window outperforms a vast window left to chance. The window is a desk; summaries are how you keep the right papers on it.
Stories persist where facts are refreshed, summaries travel, and scenes open grounded. Practise those three and the window stops being a limit and becomes an instrument.
Start the summary habit with your very next chapter break. Your longest story will be built on that small decision.
Teach the habit to anyone you story with or mentor on the platform. Summary-keeping looks like clerical work until the first time a three-month story survives an app change intact while everyone else's dissolves — after which it looks like what it is: the difference between stories as chats and stories as works.
Resist one temptation the window creates: front-loading a chat with every fact you own, in fear of forgetting. Dumping is not anchoring; it buries the present scene under the archive. Trust the rhythm instead — facts introduced as they become relevant are facts the model actually uses.
