Introduction

Previously is an AI agent that owns its memories in time. There is exactly one fundamental difference between it and every other AI assistant, and that difference changes everything: it remembers by when, not by chat thread.

No chat threads, one timeline

Mainstream assistants cut your life into disposable conversations, each starting from zero. There is no conversation list in Previously: everything you say becomes a slice on a single timeline, recording what happened and when it happened. Come back hours, days, or weeks later and you pick up the same timeline where you left it — like a long-running series where you never have to "start a new chat".

Recall is a colleague, and it answers with evidence

When you ask about the past, what answers is not a vector database but a recall colleague: it browses your timeline itself, replies in natural language, and anchors every claim to a verbatim quote. "We don't remember discussing that" is an honest, valid answer.

Meanwhile, a user card quietly updates after every conversation — who you are, what you're doing lately, what you plan to do next. It knows you not because you reintroduce yourself at the top of every chat, but because it was there.

It runs on infrastructure you own, and the memory is yours

Previously is one kernel with two ways to run it: deployed to your own Vercel project (recommended), or run locally with npm. Either way — no servers of ours, no platform database. Your memory is plain Markdown files in an ordinary git repository, yours from day one: back it up, move it, open it with any tool.

Related

• Why Previously — the full reasoning behind the idea
• Getting Started — deploy on Vercel (recommended) or local npm, running within ten minutes
• The Memory Model — slices, the timeline, and the user card in detail