Main-Context Sections

Split prompt tokens into read-only system instructions, a fixed working context block, and a FIFO message queue headed by a recursive summary.

Under OS-inspired agent memory, main-context (prompt) tokens are laid out as three contiguous sections:

  1. System instructions — read-only control flow, how each memory level is meant to be used, and how to call memory functions (including retrieving out-of-context data). This is privileged system prompt architecture, not user-editable chat.
  2. Working context — a fixed-size read/write unstructured text block, mutable only via memory function calls. In conversation, hold key user facts, preferences, and agent persona so they survive queue churn — required for engaging openers and spontaneous long-range personalization under deep memory retrieval criteria, not only for answering explicit recall questions.
  3. FIFO queue — rolling agent/user messages, system warnings, and function-call I/O. Index zero commonly holds a system message with the recursive summary of messages already evicted under memory pressure eviction.

This is a sharper cut of conversational agent context: preamble, durable working facts, and transcript are separate budgets with different write paths. Keep working context updates behind tools so the model cannot silently overwrite persona by emitting free text into the wrong section.

Sources

  • MemGPT (Packer et al.), §2.1