Memory & context

An agent’s context is the working material available to its model during a run. It can include the agent instruction, the current message, conversation history, files, tool declarations and results, and activated Skills.

Context affects behavior, latency, and cost. Memory controls how conversation history is retained as a session grows.

What enters context

Different sources enter at different times:

• Instructions define the agent’s standing goal and decision boundaries.
• Messages and history carry the current request and prior session turns.
• Tools contribute declarations that tell the model what it may call. Tool results add data after a call.
• Skills contribute short discovery metadata first. Their instructions and reference files load only when activated.
• Files include uploaded inputs and run-scoped files produced by tools.

Expose only the tools and Skills the agent needs. A larger tool surface increases declaration tokens and makes tool selection less precise. Put stable transformations behind Tasks, and use an output formatter Task when the final result needs a typed contract.

Memory strategies

Memory strategy is part of an agent revision.

max_turns is separate from memory. It limits how many runs may continue one session; memory strategy controls the history loaded when each run starts.

Compaction

Compaction can run on a cadence and under token pressure. Its configuration includes:

When a fallback model is configured, the effective context window uses the smaller supported window so a failover can continue with the retained history.

Compaction is another model operation. It can add tokens, credits, and latency, which are reported separately from foreground reasoning.

Interpret context telemetry

Agent run detail exposes context_usage:

• Provider totals report observed token usage.
• The estimated breakdown attributes input to instructions, messages, history, files, Tasks, Skills, MCP tools, Integrations, and registered tools.
• Runtime-component detail shows which resources were registered and what they contributed.
• Metadata indicates whether provider usage was complete or estimation was truncated.

The breakdown is attribution, not a context-window meter. An agent may call its model several times in one run, so cumulative run input cannot be subtracted from one model context window to calculate remaining capacity.

Observe compaction

Compaction is visible in the same evidence as the rest of a run:

• SSE and retained run events include memory_compaction events.
• phase_timings records compaction duration and trigger metadata.
• Model usage separates reasoning from memory_compaction.
• Token and credit reports show the operational cost over time.

Use Observability & audit to inspect this evidence, and Run agents via API for the event contract.

Related