# How to manage context for AI agents

> Manage context for AI agents by capturing durable work context, linking it to sources, scoping who can read it, redacting sensitive details, testing retrieval, resolving stale or conflicting memories, and auditing which agents used the context.

The right context packet is smaller than a transcript and more accountable than a summary.

## Start with durable context

Do not dump every source into a prompt. Start with the context that changes future work: decisions, conventions, skills, patterns, failures, and references.

Each memory should be concise enough for an agent to act on and specific enough for a human to inspect later.

## Attach control metadata

Good context management attaches source, scope, type, redaction state, owner, freshness, and audit history. These controls keep context useful after the original session or source artifact is gone.

A memory that cannot name its source or scope is not ready to become company context.

## Operate the loop

Run regular MemoryOps reviews: retrieval misses, boundary bleed, stale memories, unresolved conflicts, and repeated workflows that should become skills.

Ambience makes this loop visible through context readiness, scorecards, review queues, source-linked memories, and evidence of reuse.

## Related

- [Context management for AI agents](https://ambience.sh/company-context/context-management-for-ai-agents)
- [ContextOps](https://ambience.sh/company-context/contextops)
- [How do teams share AI agent memory?](https://ambience.sh/answers/how-do-teams-share-ai-agent-memory)
- [Company Context Map](https://ambience.sh/company-context/company-context-map)