A shared-memory server for AI agents

Agents are useful in short bursts and forgetful over longer ones. This sample post sketches one way to give several processes a shared notebook without pretending that every stored statement is timeless truth.

Records, not a mystery blob

The simplest useful memory is a collection of typed records. Each record carries the value itself, where it came from, when it was written, and when it should be reconsidered.

• provenance makes a fact inspectable;
• a namespace limits accidental collisions;
• a time-to-live keeps temporary observations temporary;
• plain text search provides a dependable baseline.

Semantic retrieval can sit on top later. It should not be the only way to understand why a result appeared.

Boring storage is a feature

SQLite is a good fit for the first version: one file, transactions, familiar queries, and no service to babysit. The interesting work is not the database. It is deciding what deserves to become memory and what should remain disposable context.

This post is dummy content for the site build. A proper project note will replace it once the API and conflict model settle.