Context Management
Context management is the ongoing practice of controlling what information an AI model or agent works from over time: what gets stored, what gets loaded for each task, what gets removed when it is stale, and what gets updated when decisions change. Its goal is that every step runs on relevant, accurate and current information, without overflowing the model's limited context window.
How Context Management Works
Context management operates at two levels.
Inside a session, it controls what occupies the context window. Old tool outputs that are no longer needed are cleared. Long conversations are summarized through context compaction. Large files are referenced by path and loaded only when required, instead of being pasted in up front.
Across sessions, it decides what persists. Durable knowledge, such as product rules, architecture decisions and coding conventions, is written to storage outside the window: project instruction files, decision logs, specifications or agent memory. At the start of each task, the relevant parts are retrieved and loaded back in.
A working routine usually includes:
- Store: record decisions and rules once, in a known place.
- Select: load only what the current task needs.
- Prune: remove or summarize material that no longer helps.
- Refresh: update stored context when a decision changes, so old versions stop circulating.
Anthropic, for example, describes its platform's context management features as two capabilities: context editing, which clears stale tool calls and results as the window fills, and a memory tool that stores information outside the window across sessions.
Why Context Management Matters
An AI does not know what changed last week unless someone or something tells it. Without deliberate management, two failures appear. The window fills with noise and the model loses track of what matters. Or the model receives outdated material, such as a superseded spec, and confidently builds the wrong thing.
For product teams working with AI, context management is what keeps the product's intent attached to the work. It decides whether a coding agent sees the current acceptance criteria or last month's version, and whether a pricing decision made in a planning meeting ever reaches the code.
Context Management Example
A small team building a booking app keeps its product rules in a single requirements file and its decisions in a short log. When they change the cancellation window from 24 to 48 hours, they update both and mark the old rule as superseded.
The next morning, a developer asks a coding agent to add a cancellation reminder email. The team's setup loads the current cancellation rule and the relevant story into the agent's context, but not the full product history or unrelated features. Midway through, the agent's tool clears old test outputs to free space. The agent builds the reminder for 48 hours, because the 24-hour rule never reached it. A deeper walkthrough of storing, selecting and verifying context for agents is in context management for AI coding agents.
Context Management vs. Context Engineering
The terms overlap and are often used interchangeably. Context engineering is usually framed as the design discipline: deciding what the ideal set of information is for a model at each step and building the system that assembles it. Context management is the operational side: keeping that information stored, pruned and up to date as work continues. One designs the pipeline; the other keeps it clean.