Agent context engineering

Agent Context Engineering

Produce a Context and Memory Contract that makes useful information available without letting stale, untrusted, or excessive context silently control behavior. Read references/context-and-memory-methods.md for memory classes, retrieval controls, and research sources.

Method

1. Inventory each information class and owner: current task state, source evidence, decisions, episodic history, semantic knowledge, procedures, user preferences, secrets, and derived summaries.
2. Define the context assembly order and authority rules. Preserve the distinction between source facts, retrieved memory, inference, and generated summary.
3. Assign memory class, scope, retention, provenance, freshness, sensitivity, invalidation trigger, and deletion path to every durable item.
4. Set a retrieval budget and rank by task relevance, authority, freshness, expected decision value, diversity, and token cost. Require abstention when retrieval confidence is inadequate.
5. Compile only the minimum sufficient working context. Deduplicate repeated instructions and keep conflicting evidence visible rather than blending it.
6. Treat summarization and compaction as lossy transformations. Retain source locators, material uncertainty, unresolved commitments, and a way to recover omitted detail.
7. Define handoff as a typed state transfer, not a transcript dump. Bind the recipient to objective, current state, decisions, evidence, gaps, authority, and next safe action.
8. Test stale memory, poisoned retrieval, conflicting sources, missing context, privacy deletion, compaction loss, cross-tenant leakage, and recovery after interruption.

Output

Create a Context and Memory Contract with:

• information classes, sources, owners, and authority order;
• working, episodic, semantic, and procedural memory responsibilities;
• write, retrieve, rank, compile, summarize, invalidate, and delete rules;
• token, latency, storage, and retrieval budgets;
• provenance, freshness, confidence, and conflict representation;
• handoff schema and recovery path;
• privacy, tenant, credential, and least-disclosure boundaries;
• quality and safety metrics; and
• adversarial and degradation tests.

Design rules

• Keep source-of-truth data outside the context window; context is a bounded projection with lineage, not the authority itself.
• Never make an agent-generated summary the only durable copy of a fact, decision, or commitment.
• Do not equate vector similarity with relevance, correctness, authority, or permission.
• Prefer explicit invalidation and freshness checks over indefinitely accumulating memory.
• Separate memory useful for personalization from memory necessary for task correctness, and give each an appropriate consent and deletion contract.
• Measure decision utility and error, not retrieval volume or context length.

Boundaries

• Use prompt-architecture for instruction hierarchy, skill metadata, and runtime constitutions.
• Use agent-planning-system-review for goals, plans, actions, and replanning.
• Use work-coordination-standard for portable claim, lease, checkpoint, and handoff semantics; use the matching product adapter for live operations.
• Use a product privacy or security skill for jurisdictional and threat-model requirements beyond this context contract.