Real-Time Memory Systems: How AXIOM Learns Without Forgetting

Most AI assistants have no memory. Each conversation starts from scratch. The ones that do have memory typically store it in opaque vector databases that users cannot inspect, edit, or delete. AXIOM takes a fundamentally different approach: memory is stored in human-readable Markdown files that employees can open, read, modify, and version-control with Git. The architecture maintains a semantic understanding of each employee's relationships, projects, decisions, and preferences. When AXIOM triages an email from a contact, it knows the history: previous meetings, shared projects, communication preferences, and outstanding commitments. When it prepares a meeting brief, it pulls context from email threads, task lists, and previous meeting notes — all connected through entity relationships in the memory graph. The key engineering challenge is multi-tenant isolation: in server mode, PostgreSQL row-level security ensures that one user's memories are completely invisible to another user, even if there is a bug in application code.