A personal AI agent with memory.
Ari remembers every conversation, every project, every goal I've set — for months. Reads my calendar and email with permission, drafts outbound, and tells me when I'm about to drop a ball. Runs locally on my Mac.
The "second brain" version that actually works.
Most "AI second brain" tools are notebooks with a chat interface. Ari has tools, a persistent memory file, and explicit goals it refuses to let me forget.
Reads my calendar
Scans events each morning, flags low-value meetings, drafts decline copy when I ask.
Triages inbox
Scans unread, surfaces the 3 that actually need me. Drafts replies when I say "handle it."
Tracks open goals
Persistent file of what I said I'd do, and when. Won't let me silently drop a thread.
Remembers context
Every person, project, and past conversation is in Ari's memory — for months, not sessions.
Drafts outbound
Cold reaches, follow-ups, intros. I approve before send. Ari's draft beats my blank page.
Proactive nudges
"Jon hasn't replied in 9 days" — Ari surfaces what I'd otherwise forget.
OpenClaw + gpt-5.4 + a Mac.
Nothing exotic. The trick is the persistent memory architecture — Ari writes to structured files between sessions, not a vector database.
OpenClaw — open-source persistent-memory agent runtime. Session checkpoints, tool registry, identity/soul files.gpt-5.4 via OpenAI API. Cost is pre-computed per message and logged to session JSONL.MEMORY.md, GOAL.md, IDENTITY.md, SOUL.md, DREAMS.md) that Ari reads at session start and appends to on key events.Thinking of building one?
OpenClaw is open. The hard part isn't the framework — it's the memory layer design.
See the whole fleet →