Your AI agent remembers too much. Loci for Hermes packs only 12 facts

Loci is an MIT-licensed, SQLite-backed long-term memory layer for AI agents. It ships with a generic MCP adapter and a Hermes plugin.

It sorts memories into eight types and four importance levels. Instead of injecting an endlessly growing history, it sends only 12 ranked facts into the model’s context. Ranking weighs importance, deadline proximity, and recency.

Memories also clean themselves up:

• Temporary facts expire after 24 hours.
• Daily consolidation merges duplicates and lets unconfirmed information decay.
• Completed goals expire after 90 days.

The Hermes integration can extract memories after model calls and inject relevant facts before them. The MCP adapter stays intentional and tool-driven, through remember, recall, and forget.

The repository reports 92 tests against a real temporary SQLite store.

To try it, use a disposable Hermes profile. Seed a preference, a deadline, and a temporary fact. Then inspect recall ranking, expiration, deletion, where the SQLite database lives, and exactly what the extraction hook stores.

The catch: this is a very early, single-maintainer project. Test it before trusting it with a persistent setup.