Fundamentals

RAG and Cached Context

Neutron AI supports two complementary patterns for agent memory:

Retrieval: find relevant memory for a task.
Cached context: reuse a safe, bounded context pack when a similar task repeats.

This helps agents stay consistent without sending every historical interaction back to a model.

Retrieval

Use retrieval when the task is specific, new, or likely to depend on recent changes.

const results = await client.recall({
  scopeId: process.env.NEUTRON_SCOPE_ID!,
  query: "customer onboarding constraints",
  limit: 8
});

Cached Context

Use cached context when the same agent performs repeated work for the same scope.

const context_pack = await client.agentContext({
  scopeIds: [process.env.NEUTRON_SCOPE_ID!, process.env.NEUTRON_PROJECT_SCOPE_ID!],
  agentId: "agent_success",
  task: "Prepare the next onboarding follow-up",
  cachePolicy: {
    mode: "prefer_cache",
    keyMode: "intent_profile",
    ttlSeconds: 300
  }
});

Choosing a Policy

Policy	Use when
`bypass_cache`	The task is sensitive, one-off, or must use the freshest possible memory.
`prefer_cache`	The task is repeated and can safely reuse recent context.
`refresh_cache`	You want to rebuild the context pack after a known update.

Safe Defaults

Do not cache secrets, credentials, raw payment data, or unnecessary personal data. Keep task prompts narrow and scope IDs explicit so context packs stay relevant and auditable.

Compression

Context cache