Re-inserted near the end of the conversation on every turn, not at the top. The system prompt is thousands of tokens behind by turn 30 and the model stops weighting it — this sits close to where the reply is written, so it cannot be crowded out. Keep it to two or three lines.

How many messages back from the end it goes. 0 places it after your message — strongest, but the model may answer it instead of the story. 3 is the usual choice.

Generation halts if the model produces one. One per line. Each is matched at the start of a line, so type Moodm: to stop the model speaking as you — the line break is added for you, which is what keeps it from matching the same word mid-sentence.

Speaker labels such as User: at the start of a line, and leaked ChatML control tokens. The console shows the armed list under stops. If replies ever end early for no clear reason, clear this box first — Ollama reports a stop-sequence hit and a natural ending identically, so this switch is the only way to tell them apart.

Fix it and the same prompt gives the same reply, so a change you make is attributable to the setting rather than to chance — as long as the model stays loaded exactly as it is. Measured 2026-08-17: back-to-back runs reproduce whether the model is fully on the GPU or partly on the CPU, but anything that makes Ollama RELOAD it — switching model, changing Context window, leaving it idle until it unloads — can change the answer for the same seed, because the layers are split differently the next time and the arithmetic lands in a different order. So do your A/B runs back to back, and re-take the baseline after any model switch.

Scales the cutoff to the model’s confidence instead of a fixed probability mass, so it stays coherent at the higher temperatures creative writing wants. Try 0.05 with Top P relaxed.

How far back the penalty can see, in tokens. The default 64 is why a paragraph repeated 400 tokens ago survives it. For long stories try 256–512.

Gentler than the repeat penalty, which at high values starts damaging grammar — the model avoids common words because they keep recurring.

Tokens pinned at the front — where the system prompt and the character card live — so a very long reply cannot push them out as it runs. It is not what protects you from a full window. Measured 2026-08-17: this Ollama does not trim an over-long conversation, it refuses it, with request (1078 tokens) exceeds the available context. So a chat that outgrows the window stops with an error rather than quietly forgetting who the character is — raise Context window, or start a new chat.