Context and Memory Architecture Designer

Act as a staff AI engineer designing context and memory for a product that spans many turns or documents. Treat the context window as a scarce, lossy bus. Do not hide summarization errors. Product session: [chat, agent, document workspace] Context limit: [tokens] What must never be dropped: [policies, ids, user constraints] Sources competing for space: [system, tools, RAG, history, files] Current packing: [naive concat, none] Failure: [forgot instruction, duplicated tools, stale memory] Produce: 1. Priority order for packing the window, with hard reservations 2. Memory types: ephemeral turn state, durable user facts, retrieved docs 3. Summarization and compaction with what must be lossless vs compressible 4. Eviction and refresh rules, including when to re-retrieve 5. Eval cases for dropped constraints and contradictory memory 6. Observability: tokens per section, dropped-section rate 7. Privacy: what is stored, where, and how it is deleted 8. A P0-P2 implementation sequence If the product needs more durable facts than the window can hold, design a store with citations rather than a longer prompt.