Engram, a memory-science learning tool, arrives for Hermes users
Engram is a tutoring plugin that runs alongside Hermes, but instead of helping the agent, it teaches the human using it. Its creator noticed that coding let them build systems roughly ten times faster, yet a week later they often couldn't re-explain how those systems worked — that gap motivated the tool. Engram breaks a topic into a graph of first-principles concepts and teaches one concept at a time.
Before giving any explanation, it requires the learner to attempt a guess first, following a well-established learning-science finding that retrieving an answer before being told it improves retention more than being told upfront. A separate assessor, which never sees the tutoring conversation, grades the learner's recall, so the tutor can't inflate its own results. Review sessions are scheduled using FSRS, a method that times reviews to land right before the material would otherwise be forgotten, requiring only a few minutes a day.
Wrong answers are logged word-for-word and revisited later — the creator noted having ten separate wrong-answer entries just for the concept of s.
Key points
- A tutoring plugin aimed at teaching the human user, not the AI agent
- Splits topics into a and teaches one node at a time
- Requires a guess before any explanation, based on -before-instruction learning research
- A separate blind assessor grades recall so the tutor can't inflate its own scores
- Uses FSRS to time reviews just before forgetting, and logs wrong answers for later re-testing