Model forgets earlier messages or long documents

Symptom

Long prompts are partly ignored; the model acts as if the beginning of the chat never happened.

Likely causes

Default context size is small (common in Ollama)
Silent truncation or context shifting
Parallel slots divide the context

How to fix it

Set the context explicitly (num_ctx in Ollama, -c in llama.cpp).
In llama-server, remember -c is split across -np slots.
Count tokens to confirm the prompt fits.
Use RAG or summarization instead of stuffing everything in.

bash

# Modelfile
FROM llama3.1:8b
PARAMETER num_ctx 16384

Still stuck? Check the lessons or measure first with the benchmark lab.