Use a tiered context strategy with explicit context budgeting: Task Context → Candidate Memory Retrieval → Relevance Ranking → Priority Filtering → Compression → Context Assembly Separate memory into active context, working memory, episodic memory, semantic memory, procedural memory, and external knowledge. For each task, allocate the context budget across system instructions, current requirements, workflow state, retrieved memory, and external evidence. Do not inject the complete conversation by default. Preserve decisions, constraints, dependencies, identifiers, and authoritative evidence. Compress repetitive history while retaining a reference to the original artifact. Track task_id, workflow_state, completed_steps, pending_steps, decisions, and artifacts as structured state instead of reconstructing critical state from conversation history. Evaluate the design using context-token usage, irrelevant-context rate, retrieval precision, task completion, latency, and answer quality.