Choose what context to keep

Use typed questions before trimming a long agent history.

Judge relevance one unit at a time

When a conversation grows, split it into candidate facts or tasks. Ask whether each candidate is still relevant to the next action. Keep source pointers so the agent can recover detail later.

Protect non-negotiable context

System instructions, user constraints, and unresolved commitments should be retained by code rules. A model can suggest compression candidates, while the host decides what is mandatory.

Open the workbench Next: Jev or a general LLM?

JevModel is independent and not affiliated with TypeSafe AI.