keep_alive: -1 — keeping an Ollama model in RAM between calls

By default, Ollama unloads a model from RAM after 5 minutes of inactivity.

The next request reloads it from disk. On CPU, that takes 10-15 seconds.

For an agent generating messages on demand, that's a visible delay on every call spaced more than 5 minutes apart.

The fix

Pass keep_alive: -1 on every call.

await ollama.chat({
  model: "qwen2.5:7b",
  messages,
  stream: true,
  keep_alive: -1,          // keeps the model in RAM indefinitely
  options: { num_predict: 100 },
});

The model stays loaded between calls. No more 13-second reloads.

When not to use it

If the machine has little RAM and several models are loaded in parallel, keep_alive: -1 can waste memory unnecessarily.

The default 5-minute value is reasonable for interactive use (one-off questions). It becomes a penalty for an agent generating messages at irregular intervals — sometimes 2 minutes apart, sometimes 20.

Centralizing the client

Instead of passing keep_alive on every call in every file, use a single shared Ollama instance:

// src/ollama/manager.ts
export const ollamaClient = new Ollama({
  host: config.ollama.host,
  fetch: fetchWithTimeout(10 * 60 * 1000),
});

And keep_alive: -1 on each chat() call. Every module imports ollamaClient instead of instantiating its own new Ollama().