Context Management
Strongly agents handle arbitrarily long conversations without exceeding the model's context limit. You pick the strategy at build time and adjust later in Agent Settings → Context Policy.
Three policies, plus two knobs. That's the whole set.
Pick a policy
Window
Keep the most recent turns; drop older turns once the budget is hit. This is the platform default.
- Cost: lowest.
- Recall: lowest. The agent won't remember turns it dropped.
- Pick this for: short conversational agents, support triage, one-shot Q&A.
Summarise
Keep a rolling summary of older turns plus the most recent turns verbatim.
- Cost: medium -- one summarisation pass when the budget tightens.
- Recall: good. Decisions made earlier survive into the summary.
- Pick this for: personal assistants, long threads, anything where earlier context matters later. This is Iris's default.
Hierarchical
Three tiers -- recent turns verbatim, older turns summarised, oldest turns reduced to titles.
- Cost: highest.
- Recall: best.
- Pick this for: research assistants and multi-day operational threads.
The two knobs
| Knob | What it does | Sensible default |
|---|---|---|
| Token budget | Max tokens of conversation history kept after applying the policy. | 70% of your model's context window |
| Keep recent turns | The minimum number of recent turns retained verbatim, regardless of policy. | 12-20 |
The wizard auto-fills the token budget based on the primary model you picked. You can override it, up to the model's context window: a larger budget is marked and cannot be saved, and switching to a model whose window is smaller than the budget asks you to lower the budget first.
When edits take effect
A saved context policy reaches a running agent when you click Apply changes in the Settings header, which stays highlighted until you do.