Self-Driving AgentsGitHub →
Powered by Hindsight · Portable agent memory

Agents that learn from every conversation and get better over time.

Install in one command. They build their own knowledge pages, remember what you like, and keep themselves current — across Claude Code, Claude Chat, OpenClaw, and NemoClaw.

npx @vectorize-io/self-driving-agents install marketing/seo --harness claude-code

Replace the agent slug or harness flag with whichever combo you want.

How it works

You never tell the agent to "save" or "remember." It decides what matters and keeps itself current.

1

You chat with the agent

Use it like any normal assistant. Ask questions, give feedback, share context.

2

Memory is retained

Conversations are auto-summarized into Hindsight as durable, structured memory.

3

Knowledge pages update themselves

After each session the agent rewrites the pages it owns — playbooks, references, preferences.

4

Next session it remembers

On the next chat the agent reads its current pages — it knows what works and what you prefer.

Want the deeper picture?

See how the harness, the agent, and Hindsight fit together — what crosses the boundary, what stays server-side, and why pages and memories live in two layers.

Available agents

Install the whole department or pick a specialty.

See all agents →

Or roll your own agent.

An agent is just a directory: drop a bank-template.json next to some markdown files and point the CLI at it. Same self-driving behavior, your domain.

Read the spec →
my-agent/
  bank-template.json     # memory bank + knowledge pages config
  playbook.md            # any .md/.txt becomes seed knowledge
  advanced-tips.md
npx @vectorize-io/self-driving-agents install ./my-agent --harness claude-code