4 min read
Context Windows Are Not Product Memory
LLMsArchitectureAI

Cover art for LLM context and memory design
Stuffing “everything important” into the prompt is a temporary patch, not architecture.
Prefer external memory: ticket systems, docs, user profiles, and vector stores with clear ownership of what can be recalled.
Summarize for sessions, store for durable facts. Treat auto-summaries as lossy caches, not source of truth.
Be deliberate about what never enters the model context: secrets, sensitive identifiers, or data the user did not intend to share.
When the model contradicts stored facts, the product should trust storage and show the user what changed.