All writing
2026-07-28·3 min
Memory shouldn't be treated as context
Stuffing history into a prompt is a hack, not an architecture.
There is a popular pattern: take the last N messages, embed them, retrieve similar chunks, and prepend everything to the prompt. This is context management, not memory.
Memory implies persistence, structure, forgetting, and consolidation. Context is just what fits in the window right now. Treating memory as context makes agents stateless in disguise, which means they cannot learn, plan, or build trust over time.
We need to separate the working context from the long-term memory store. The working context is small, ephemeral, and task-specific. The memory store is large, durable, and queryable. The interesting engineering is in the boundary between them.