Agent Memory, Follow-ups & Scheduled Runs

The difference between a chatbot and a colleague is that a colleague remembers, chases, and shows up without being asked. These are the three primitives that make the AI agent a standing operation rather than a one-shot Q&A box.

Persistent memory

Tell the agent something once and it holds onto it. Memories are workspace-scoped facts — a decision, a commitment, a preference, or context — and the agent recalls them before answering. Say "remember that we bill Acme monthly, not per-project" and it will still know next week. Memories can carry an expiry, and you can ask it to forget one.

The waiting-on ledger

Work stalls on things other people owe you. The agent keeps a ledger of who owes what by when — ask "what am I waiting on?" to see it. It records each time someone is nudged, and after a couple of nudges it will tell you to escalate rather than send a third.

Scheduled autonomous runs

A schedule is a recurring prompt the agent executes on its own — the "email me every morning what I should work on" case. Create, edit, pause and read them on Sidebar → Workspace → Scheduled Runs (/ai-agent/runs) — start from a template or Write a custom run — or just ask the agent to set one up. Any member can create one and manage their own; owners and admins see every schedule in the workspace, there and in Settings → AI Agent.

Schedules run in UTC. The UI shows you the local-time equivalent as you pick, so 8am Pacific becomes 16:00 UTC without you doing the arithmetic.
Safeautonomy (the default) gives the agent its full tool loop with external sends, deletes, and payments stripped out — it can read and reason, but it can't email your customers; its own digest to you still goes out. None makes the run a single text reply.
Run as binds a schedule to one of your AI employees, and Only send when skips a run when its person has no tasks, so an empty digest is never sent.
Every run injects the agent's newest memories and the open waiting-on ledger, so a scheduled briefing knows what you're blocked on.
Run history — status, summary, any error, and the credits each run cost — is on the Scheduled Runs page. Owners and admins can open See what Kavanah skipped today, and why there for anything held back in the last 24 hours.

Runbooks and goals

A runbook is an ordered SOP the agent executes step by step, pausing at checkpoints until a human confirms. A goalis a metric with a target and a threshold — the agent evaluates it daily and only raises an alert when the verdict gets worse, so a chronically off-track goal doesn't nag you every morning.

All of this is reachable by just talking to the agent — "remember X", "what am I waiting on?", "brief me every weekday at 8". You never have to learn a command syntax.

Persistent memory

Scheduled autonomous runs

Schedules run in UTC. The UI shows you the local-time equivalent as you pick, so 8am Pacific becomes 16:00 UTC without you doing the arithmetic.

Safeautonomy (the default) gives the agent its full tool loop with external sends, deletes, and payments stripped out — it can read and reason, but it can't email your customers; its own digest to you still goes out. None makes the run a single text reply.

Run as binds a schedule to one of your AI employees, and Only send when skips a run when its person has no tasks, so an empty digest is never sent.

Every run injects the agent's newest memories and the open waiting-on ledger, so a scheduled briefing knows what you're blocked on.

Run history — status, summary, any error, and the credits each run cost — is on the Scheduled Runs page. Owners and admins can open See what Kavanah skipped today, and why there for anything held back in the last 24 hours.

Runbooks and goals

All of this is reachable by just talking to the agent — "remember X", "what am I waiting on?", "brief me every weekday at 8". You never have to learn a command syntax.