# Context

The selected information available to an agent during a current decision — the objective, the policy, the conversation so far, the business records, the retrieved material and the working notes.

Source: https://aecon.ai/glossary/context
Updated: 2026-09-15

## In practice

Before drafting a refund response, the system assembles the customer's order, the current refund policy version, the prior conversation and the agent's spending limits.

## The distinction

A large context window is temporary capacity, not knowledge. Context determines what the system can consider now; it does not establish that any of it is current, permitted or true.

## Seven kinds of information, seven different controls

| Component | What it tells the agent | Primary risk |
| --- | --- | --- |
| Objective | The result and the stopping condition | Ambiguous or conflicting goals |
| Policy and authority | What it may do, spend, disclose or change | Instructions mistaken for permission |
| Current interaction | What has happened in this run | Constraints displaced by noise |
| Business state | The current customer, order or account record | A stale copy treated as the system of record |
| Retrieved knowledge | Relevant policies, research or prior cases | Irrelevant, untrusted or unauthorised records |
| Retained memory | Selected state from earlier sessions | False, sensitive or obsolete memories |
| Working material | Plans, calculations and temporary files | Leakage or uncontrolled retention |

Policy normally outranks a user preference. A current account balance comes from the banking system, not from a memory written last week. Keeping the categories apart is what makes those rules enforceable.

## Retrieved content is data, not command

Web pages, emails and documents can contain text that reads like an instruction to a model. When an agent can act, a hostile or accidental instruction inside retrieved content can attempt to redirect the task or invoke a tool.

The operating hierarchy places policy and explicit delegated authority above retrieved content. Untrusted material stays labelled, tool permissions stay narrow, and consequential actions require approval.

## Sources & context

- [Manage tool context](https://platform.claude.com/docs/en/agents-and-tools/tool-use/manage-tool-context) — Anthropic, 2026. Documents prompt caching, on-demand tool definitions, programmatic tool calling and context editing as ways to manage context pressure. Checked 2026-09-13.
- [Retrieval](https://platform.openai.com/docs/guides/retrieval) — OpenAI, 2026. Documents semantic search over vector stores with attribute filtering applied before search, and results carrying the originating file. Checked 2026-09-13.

## Related reading

- [Memory](https://aecon.ai/glossary/memory)
- [Retrieval](https://aecon.ai/glossary/retrieval)
- [Agent harness](https://aecon.ai/glossary/agent-harness)
- [Provenance](https://aecon.ai/glossary/provenance)