Skip to main content
Each model request draws from one agent’s perspective. The runtime may host several agents, but one agent’s private conversations, identities, settings, and memories do not become another agent’s background knowledge.

​
What enters a turn

The history-search tool can retrieve older or other-conversation messages the current audience may receive. It does not bypass the same permission rules used for recent history.

​
Two independent boundaries

Perspective answers whose observations may be considered. All graph and memory reads are keyed by the current agent. A deployment-wide identity match or another agent’s access to a room grants nothing here. Audience answers where an observation may be reused. A private memory does not become channel-visible because it is relevant or because its subject joins the channel. Memory privacy describes authenticated source evidence, derived lineage, sharing grants, and the checks before delivery. Verified identity links allow one person’s eligible private history to follow them between direct conversations within this agent’s perspective. They do not merge agents or broaden group permissions.

​
Prompt assembly

The initial request separates runtime-authored system instructions from attributed external context and the current message. External text does not become system instructions simply because it arrived in a memory or transcript. Prompt admission reserves required content and tool schemas, then fits whole memory and history records into the remaining capacity. Optional records may be omitted; required content that cannot fit causes a budget error before provider execution. Context already exposed to a model remains in the turn’s provenance even if later calls omit it. Changed assembled prompts are saved before the model call when snapshot storage is configured. Model-call records refer to those snapshots. Checkpoint recovery rebuilds live system policy and current permissions; saved system instructions are not replayed as authority.

​
Implementation

Recent history lives in internal/graph/history.go (repository), memory selection in internal/memory (repository), and prompt admission in internal/prompt (repository). The turn engine combines them. There is no separate shipped WorkGraph worker context resolver or cross-agent shared-memory pool.