Remember: A Per-Folder Protocol for Personal AI Memory

Remember is a protocol for personal AI memory that stores all data as plain markdown on the user's own filesystem, scoped per folder rather than per user. It specifies four primitive verbs (Remember, Ingest, Refresh-wiki, Lint) and a flagship application (Checkin); a four-layer storage model (a raw journal, typed atoms, a deterministically regenerated index, and an optional synthesized wiki); a continuous active maintenance design in which a per-response Stop hook writes every turn to the journal (a structural guarantee on capture) while extraction into typed atoms is a separate, explicit, audited step; and an anti-fabrication ("gate-set") discipline that verifies every user-facing memory claim against disk state before surfacing it. The paper positions Remember against eight contemporary memory systems and is explicit about where peers do specific things better. Status: This is a design-and-protocol paper with a reference implementation. Its empirical claims (capture coverage, cross-context contamination, index stability) are argued from design experience and small-scale observation; systematic evaluation is planned for a future version. Feedback and critique are actively welcomed - please contact the author. Reference implementation (MIT-licensed): https://github.com/sujayathali/remember