# Summarising Old Turns — Agent Loops, Stop Conditions and Token Budgets

Source: https://www.geekswithgeeks.com/en/agent-loops/ctx-summarise

> Replace old messages with a faithful summary and keep the task, decisions and open items.

## Compress, do not erase

When history passes your trigger, ask the model (or a cheaper one) to summarise older turns into a short note, then replace those messages with it and keep the **most recent turns verbatim**. A good summary preserves the goal, decisions made, files changed and what is still to do. Facts the summary drops are gone, so keep exact values (IDs, paths) in it.

## Summarise, keep the tail

`summarise` is another model call. Never summarise away the first user message, because it holds the original task.

```python
def compact(messages, keep_last=6):
    if len(messages) <= keep_last + 1:
        return messages
    head, old, tail = messages[0], messages[1:-keep_last], messages[-keep_last:]
    note = summarise(old)       # goal, decisions, open items, exact IDs
    return [head,
            {"role": "user", "content": f"Summary of earlier work:\n{note}"},
            *tail]
```

## Check the summary

Summaries can invent or drop details. For important runs, log the original messages somewhere before replacing them so you can audit what was lost.

**Quiz:** What should never be summarised away?

- [x] The original task
- [ ] Greetings
- [ ] Old error logs
- [ ] Whitespace

*Answer:* The original task. Losing the original goal makes the rest of the work aimless.