The assistant in my pocket that can no longer crash its own host
Trig, the AI assistant I built, runs on a Mac Mini in another room. This puts it in my pocket: a Slack DM or mention becomes a conversation with the same memory my desk sessions use.
Each thread is its own persistent session, keyed in SQLite so it survives restarts and a thread still feels like one continuing chat days later. Memory is scoped by who is asking. I get the full context; my VA, Carla, gets only her own operating file, with no personal or strategic detail; anyone unrecognised is dropped before a reply is ever generated.
Consequential actions sit behind a two-layer gate. The first layer is a hard block that stops the bot ever issuing a command to restart its own host process, and it cannot be overridden by any approval. The second requires an explicit approval phrase before a state-changing action runs, while read-only checks pass through freely. The first layer exists because the bot once did restart itself mid-reply; it is now structurally unable to.
