Memory, context and the story summary

AI models can only read a limited amount of text at once. RPGCortex uses several layers to make sure the most important parts of your story always fit.

What the AI reads each turn#

Every turn, the AI receives a carefully assembled package. In order, it contains:

1. The narrator's base instructions.
2. The scenario's Plot Essentials, content guidance, AI Instructions and Story Background.
3. State Cards — the always-on ones, plus any whose trigger keywords appear in recent turns (see State Cards).
4. The Story so far summary.
5. Relevant memories from the Memory Bank.
6. The current game state — stats, inventory, characters present and so on.
7. The Author's Note, deliberately placed near the end so it has a strong effect on the writing style.
8. Up to the last 10 turns of the story.
9. Your new action.

Context: how much fits#

The size of that package is measured in tokens (roughly four characters each). Each plan includes an amount of free context per turn:

Each model also has its own maximum, which can be lower than your plan's amount.

If you have credits, turns can automatically go beyond your free context — each extra 4,000 tokens (or part of it) costs 1 credit. Some models charge for every token instead. Both are explained in AI models. On standard models, if you're out of credits, turns simply stay within your free context — you're never blocked.

When the story doesn't fit#

If everything won't fit, RPGCortex trims in this order:

1. The oldest turns of history, one whole turn at a time.
2. Then memories, least relevant first.
3. Then keyword-triggered State Cards.
4. Then the story summary.

Always-on cards, the scenario's instructions and your current game state are never trimmed. If those alone are too big, you'll see a warning explaining which parts are largest — shorten them, or switch to a model or plan with more context.

The context meter#

Open Settings → Gameplay → Models → Context Usage to see exactly how this turn's context is being spent. The colored bar breaks it down into parts such as System instructions, Setting (plot essentials), AI Instructions, Author's note, Story background, Story so far (summary), State cards (triggered), Memories, Stats, Inventory, Characters, Recent turns and Your input. It updates as you type.

The Context length slider#

Under the meter, Context length controls how many tokens are sent each turn.

• By default it's set as high as you can currently afford.
• Lower it for faster responses, or to avoid spending credits — Reset sets it back to your plan's free amount.
• The help text tells you whether this turn would cost credits, and how many.
• The setting is remembered for each adventure on this device.

Response Length#

Also under Models, Response Length caps how long each AI response can be, from 50 to 500 tokens (500 by default). Shorter responses are faster.

The Memory Bank#

Each turn — your input and the AI's response — is saved as a memory. On later turns, RPGCortex searches your recent memories for the ones most related to what's happening now and includes up to five of them. That's how a detail from fifty turns ago can resurface at the right moment.

How far back it searches depends on the host's plan:

Upgrading takes effect on your existing adventures straight away. Memories only come from the branch you're on, and editing a passage updates its memory.

The story summary#

As your adventure grows, RPGCortex keeps a running Story so far summary, refreshed roughly every eight turns. It's what lets the AI remember the broad strokes of your tale long after older turns stop fitting.

• You can read the current summary in the context meter, under Story so far (AI-maintained, read-only).
• If you've been away for more than six hours, a Welcome back — here's where things stood banner shows the summary when you return.
• Editing an earlier passage causes the summary from that point onwards to be rebuilt.
• To turn it off, untick Auto Summarization in Settings → Scenario → Details — but be aware a long adventure may then lose track of older events.