RemindrAI: An Embeddable, Multi-Tenant Reminder and Multi-Channel Communication Capability for Conversational Applications

Conversational applications increasingly need to act on what they discuss, delivering a message later on whatever channel the user checks. This capability, reliable time based delivery to an identified person, gets rebuilt inside nearly every product that needs it, hitting the same problems each time: interpreting how people express time, delivering a message exactly once despite retries and crashes, advancing recurring schedules across daylight saving boundaries, within a serverless platform's tight execution budget. This thesis presents RemindrAI, a multi tenant reminder and communication backend built during an internship at MemorAIz S.r.l. It solves this problem once, behind a stable contract, so other applications can consume it instead of rebuilding it. Three decisions matter most. Retry authority lives in the database, not the queue: reminders are claimed atomically, delivery is idempotent per channel, and attempt state is just a column, so the guarantee holds regardless of the queue runtime. Identity authority stays separate from reasoning: a tool calling agent chooses operations and arguments, while ownership and tenancy come from the service, never the model, enforced at four depths. Channel selection happens in one step, combining organizational policy with user preference, with provider limits declared as adapter capabilities rather than scattered conditionals. The backend ships as a Mastra capability an agent can register in one statement, and runs in production for the H FARM Student Assistant, handling study reminders, account linking, and staff broadcasts. The evaluation rests on 670 automated tests and focuses on named invariants rather than performance figures, since the system has no instrumentation for latency or throughput, a gap named explicitly. Two negative findings are reported openly, both later closed. The contribution is architectural: reliable delivery, multi tenancy, and agent extensibility compose into a portable capability.