Skip to main content
Every agent turn ends in a single model request: one system prompt, the conversation history, and the tool list. This page explains which values the model can see each turn, where they come from, and how to keep a value out of the prompt entirely. For how that request is kept small — the size budgets and compaction policies — see Context size, compaction, and budgets.

​
At a glance

Before the model reads it, every source of context is shaped: memory and ## Current Context by visibility rules and size limits, conversation history and tool results by compaction, and knowledge-base results by a token budget. Guardrails check what goes into and comes out of each call, but they never see the assembled prompt itself.

​
What the model sees each turn

Session values reach the system prompt in two ways: through the memory sections an agent declares (## Session Memory, ## Granted Memory, and similar), and through a ## Current Context block. Which values land in ## Current Context depends on the agent’s visibility mode. These values are never shown in ## Current Context, in either mode:
  • Platform identifiers — session, tenant, project, and user IDs — plus raw input and message metadata.
  • The config and env namespaces.
  • Keys that start with an underscore, except the handoff summary.
  • Knowledge-base search results. They reach the model through the search tool call instead — see Knowledge base retrieval.
  • In allow-list mode, values already shown in a memory section. In explicit mode, a declared session value can appear in both a memory section and ## Current Context, so budget for that when sizing prompts.
Tool results are visible only to the agent that produced them. A last_<tool>_result value appears in the prompt of the agent that called the tool — see Store result. After a handoff, the receiving agent does not see it unless it’s listed in pass or referenced directly.

​
Size limits

Each ## Current Context value is limited to 8,000 characters by default, and the whole block to 24,000. Automatic tool-result values are limited to 4,000 characters. A value over its limit is replaced by a marker object holding a preview of its start. Declared memory sections (## Session Memory and similar) are not size-limited, so keep declared memory small. The full set of default budgets, the voice-profile overrides, and how to diagnose a bloated or truncated prompt in a trace live in Context size, compaction, and budgets.

​
Keeping a value out of the prompt

To keep a value out of what the model reads:
  • Don’t declare it in MEMORY: and don’t reference it. An allow-list agent never sees an undeclared value it doesn’t reference.
  • If a tool needs the value, declare it as a runtime-supplied parameter the model can’t see.
  • For tool results, use the tool’s MODEL_CONTEXT to pass the model only the fields it needs — see MODEL_CONTEXT on a tool result.

​
Referencing values in prompt text

Use {{customer_tier}} in GOAL, not {{context.customer_tier}}. In GOAL, PERSONA, and LIMITATIONS, the context. form makes the value visible in ## Current Context, but the placeholder itself is not filled in — only the plain {{var}} form interpolates. If a referenced value is unset, the literal {{customer_tier}} text stays in the prompt. Word the sentence so an empty value still reads sensibly, or set a default with INITIAL:.
There is no per-entry setting to hide one declared memory value from the model while leaving the rest of a MEMORY: SESSION: block visible. Use the options in Keeping a value out of the prompt above instead.
Related articles: