Technical topic
Context engineering, memory, and state
Production context assembly, memory governance, durable workflow state, provenance, and side-effect safety.
Direct answer
Context engineering decides what information a model receives for the current step, while memory governs reusable information and workflow state records durable execution truth. Production systems must give each layer its own authority, provenance, retention, versioning, and recovery rules so stale context, poisoned memory, or an ambiguous retry cannot silently change an outcome.
What this topic helps you decide
AI context engineering
Allocate limited context to task instructions, evidence, history, tools, and output constraints deliberately.
Agent memory governance
Control admission, provenance, scope, expiration, correction, and deletion across sessions.
Durable agent workflows
Persist state, executable versions, authority, attempts, and confirmed effects for safe resume and retry.
Practical questions answered
- Context, memory, and state: what is the difference?
A production design boundary for deciding what the model sees now, what the system may remember later, and what remains authoritative workflow truth.
- How do I allocate a context-window budget?
A context-frame budget for instructions, authority, state, evidence, tools, history, uncertainty, and output reserve in production agents.
- How do I prevent stale or poisoned agent memory?
A memory lifecycle with admission, provenance, scope, expiry, promotion, conflict handling, quarantine, and rollback for production agents.
- How should retrieval preserve provenance?
An evidence supply chain connecting source records, versions, chunks, retrieval runs, context placement, claims, citations, and downstream effects.
- How can a long-running agent resume safely after hours or days?
A resume contract for durable agents that separates checkpointed state, external effects, versions, leases, approvals, revalidation, and recovery.
- Why did my agent perform the same side effect twice after a timeout?
A retry-safe effect protocol using operation identity, idempotency, effect receipts, reconciliation, and explicit unknown completion states.
- What happens if approval or authority expires while an agent is waiting?
A commit-time authorization check for agents whose identity, approval, policy, target, or proposed action can change during a pause.
Go deeper with a field guide
Context Engineering in Production
Designing Memory, Retrieval, State, Provenance, and Long-Running Agent Workflows
Explore Context Engineering in ProductionRelated books
Reusable resources
- Agent run envelope schema (JSON Schema)