An agent's memory is a database URL

You give the agent a job that takes a week: research, summarize, follow up, revise. Across a hundred LLM turns, it needs memory: what it learned, what it tried, what it discarded, what's pending.

You don't pip install a vector store or a memory framework. You give the agent a SQLite KV URL. Whenever it learns something worth keeping, it PUTs it under a key. Whenever it needs to recall, it GETs by prefix and reads back the rows. Atomic incr tracks "times this question came up." TTL drops scratch notes after a day.

The agent dies and respawns on Tuesday. Same URL, same keys, same memory. A second agent joins the project: same URL, same memory, no migration. Memory becomes infrastructure, not framework state.