Support conversations are rarely short. A customer writes in, an agent asks a question, a colleague adds a note, the thread goes quiet for two days and then picks up again. Until now, every one of those turns stayed in the agent's working memory in full, and once the history grew past the model's context window the run failed.
That limit is gone. Agents now compact their own history while they work.
How it works
Every agent keeps a rolling window of recent turns verbatim. Once the history grows past a configurable threshold, the agent writes a short summary of the older turns and keeps only that summary plus the recent window. The next message is answered with the same context the agent has always had, just expressed in fewer tokens.
Three things stay true through the whole process:
- The recent turns are never touched. Only older history is summarised, so the detail of what just happened is always intact.
- The full transcript is still stored. Summarisation affects what the model sees on the next turn, not what you see. The session log keeps every message.
- Summaries are visible. Each compaction is recorded in the execution log, with the summary text, so you can always see what the agent chose to carry forward.
What you should notice
Long-running conversations simply keep working. Practically, that means:
| Before | Now |
|---|---|
| Long threads failed once history exceeded the context window | Threads keep running, history is compacted automatically |
| Cost grew with every turn in the thread | Cost stays roughly flat after the first compaction |
| Manual restarts to clear the history | No restarts needed |
Cost is the quiet win here. In a thread of a hundred messages, every turn used to re-send the whole history. With compaction the per-turn payload settles at a predictable size instead of growing forever.
Turning it on
Summarisation is on by default for new agents and can be adjusted per agent in the agent settings. You control how much recent history stays verbatim and how aggressively the older part is compacted. If you would rather keep the raw history for a specific agent, you can switch summarisation off and the agent falls back to a hard cap on history length.
If you are running an agent where every earlier detail matters, keep the verbatim window generous. Summaries are good, but the raw turn is always better than a description of it.
As always, nothing about this changes where your data lives or how it is used. Conversation content is not used to train models, by us or by any model provider we route to.