Context Summarization
Automatically summarize older messages when the context window fills up — for both CugaAgent and CugaSupervisor.
For long conversations, CUGA can roll older turns into a running summary so the LLM keeps the most useful context without blowing the window.
The full option list lives in the Settings reference — Context Summarization.
Enable
[context_summarization]
enabled = true
keep_last_n_messages = 10
trim_tokens_to_summarize = 500
summarization_model = "gpt-4o-mini"
trigger_fraction = 0.75With this configuration:
- Summarization fires when the prompt would exceed 75 % of the model's context window.
- The last 10 messages are always preserved verbatim.
- Older messages are condensed into ~500 tokens by
gpt-4o-mini.
Trigger options
You can use any combination of the three trigger conditions; whichever fires first wins.
| Trigger | Use when |
|---|---|
trigger_fraction = 0.75 | You want the trigger to track the model's actual context window — recommended for production. |
trigger_tokens = 2000 | You want a fixed token cap regardless of model. |
trigger_messages = 20 | You want to summarize after a fixed number of turns (useful for testing). |
If you set more than one, the first condition that becomes true triggers summarization.
Custom prompt
By default LangChain's built-in summarization prompt is used. To override:
[context_summarization]
custom_summary_prompt = "Provide a concise summary of the following conversation, preserving all numeric values and named entities: {messages}"The {messages} placeholder is the only required variable.
Choice of summarization model
summarization_model is independent of the agent's main model. Most users keep it on a small/cheap model (gpt-4o-mini, claude-haiku, etc.) — the goal is fast, lossy compression, not high reasoning.
Works with CugaSupervisor
Context summarization applies to both CugaAgent and CugaSupervisor runs. Each delegated sub-agent invocation gets the summarized history just like a standalone agent.
Summarization is lossy by design. If your task depends on remembering every literal detail (e.g. exact figures from a document), prefer the Knowledge Base — it keeps the original document available for retrieval.