Selective memory: keywords over transcripts

The naive way to give an assistant memory is to stuff the whole transcript back into the context window. It's expensive, it's noisy, and it gets worse with every turn. We went the other direction: remember less, but remember the right things.

Keyword memory

Instead of storing conversations verbatim, MasterNode extracts compact, structured facts — entities, preferences, and durable details — and injects only what's relevant to the current turn.

• Entity-first: facts are keyed to the things they describe.
• Preferences persist globally so 'always use metric units' survives across chats.
• Retrieval pulls a minimal, relevant slice — not the whole log.

Why it matters

Compressed memory keeps prompts small, cheap, and focused. The model stops drowning in old context and starts using the handful of facts that actually change the answer.