max_context_tokens sets an upper bound for the assembled context. The assembler tracks:
  • token_counts: planned, budget, dropped
  • budget_actions[]: why items were dropped or compressed
Tips:
  • Lower the preset top_k or increase the budget to include more facts.
  • Use must_include to pin critical memory_ids.
  • Enable return.include_structured_facts to render facts compactly in prompts.