Skip to main content

Transcript, Conversation State, and Prompt Context

Read this if: you need the exact continuity model for long-lived conversations under compaction.

Skip this if: you only need the high-level conversation model; start with Messages and Conversations.

Go deeper: Context, Compaction, and Pruning, Memory, Work board and delegated execution, and ARCH-20 conversation and turn clean-break decision.

Parent concepts​

Scope​

This page defines the three durable layers that are often confused with one another: transcript, conversation state, and prompt context. Tyrum keeps them separate so one conversation can survive many compactions without losing current truth.

Continuity stack​

Transcript​

Transcript is the retained event history for one conversation.

  • It is append-only.
  • It exists for audit, replay, troubleshooting, and operator inspection.
  • It may contain many more events than the model can ever see in one prompt.
  • It must survive compaction because compaction is not deletion of history.

Conversation state​

Conversation state is the mutable continuity layer for one conversation.

It should capture the smallest durable set of facts needed to continue safely across turns, such as:

  • compaction checkpoint or handoff summary
  • recent event cursor
  • pending approvals
  • pending external or tool activity
  • pinned current-truth values that should not drift under transcript compaction

Conversation state is not long-term memory and not a second transcript. It is the continuity projection of the conversation.

Prompt context​

Prompt context is the bounded model input for one turn.

It is assembled from:

  • system prompt and runtime instructions
  • conversation state
  • the recent transcript tail
  • memory recall
  • WorkBoard digest
  • current turn input and attachments

Prompt context is ephemeral. It can change on every turn without changing the underlying transcript or conversation identity.

Why this separation matters​

  • One conversation can survive many compactions because transcript is retained while conversation state carries forward current truth.
  • The model does not need the full transcript to continue safely.
  • Memory and WorkBoard stay outside the conversation transcript, so long-term knowledge and work status do not depend on prompt replay.

Compaction rule​

Compaction transforms older transcript history into a tighter conversation-state checkpoint plus a recent transcript tail. It does not create a new conversation and does not replace transcript retention policy.