Long Sessions

When the context window fills, Caretz summarizes older turns and keeps recent ones word for word. This can repeat any number of times, and the full history stays on disk. Check or compact an interactive session directly:

/context
/compact

Most users only need to configure the model’s real context limit. To tune the budgets, see Compaction budgets .

How Compaction Works

Caretz estimates usage against the configured context window. It reserves room for the next request and response, then summarizes older complete turns when space becomes tight.

• A later compaction carries the previous summary forward, together with the active request, current task state, and newer conversation.
• During long tool chains, Caretz can summarize completed tool-call boundaries while preserving the active user request.
• If a provider reports a smaller context limit than configured, Caretz rebuilds the summary request against that limit and retries.

This design has supported hours-long local sessions around a 60k context window during Caretz’s development. That is field experience rather than a guarantee: summary quality, tool calling, server limits, and configured budgets still matter.

Compaction never deletes the audit history. It appends summary events and changes the active model view reconstructed from context.jsonpl.

Keep Multi-Step Work on Track

Internal task tracking is enabled by default. For multi-step work, the model can record the current step, completed work, and a concrete next action. That state is persisted separately from ordinary conversation messages, restored with the session, and included as the authoritative work state when Caretz builds a compaction summary. A task list therefore survives both compaction and restart.

This is especially useful for smaller models: leave tasks enabled for long or multi-step jobs so progress survives context compression. For a narrow, single-purpose agent, remove the task tools and guidance with --no-tasks.