DoorDash built an AI that remembers what it did last week

DoorDash's engineering team built an internal AI assistant that actually works across sessions. Most chatbots forget everything once you close the tab. This one remembers. You ask it to look up a report, it does, then you say "now show me last month's numbers" — and it pulls the right context without making you repeat yourself.

The trick isn't magic. It's simple bookkeeping. The agent keeps a lightweight summary of what's happened, then fetches the full details only when it needs them. Instead of dumping the entire conversation history every time, it writes a running journal. When the journal gets too long, it compresses the older stuff into a shorter summary. The math is straightforward: less data to send to the model, same quality results.

For an engineering team that runs dozens of reports, queries databases, and debugs code all day, this matters. The assistant can pick up a task at 10am and finish it at 4pm without losing track. It's the kind of tool that saves more hours than the hype promises.

Why this matters for us: As AI gets folded into the tools we use daily, the ones that actually remember what they did are the ones that save us time — not the ones that make us repeat ourselves.

> "The assistant can pick up a task at 10am and finish it at 4pm without losing track."