How do I run multiple Claude Code agents with persistent memory?

Give each agent its own install, its own memory and its own checkpoints. In aigent-os each seat has its own memory vault and its own capsules, and the maintainers run four seats (Titus, Pheme, Davinci and Metis) on one business, The AIgent newsletter.

How aigent-os does it

• Seats. One install is one seat: one long-running Claude Code session with its own vault and capsules. Run more seats for more agents. See Creating agents.
• Specialist sub-agents. Inside a seat there are nine named agents: a builder, a read-only scout, a researcher, a design spec writer, a memory architect, a critic, a vault custodian, an architecture reviewer and a prompt writer, each with a fixed model tier and tool set.
• Model tier check. A PreToolUse hook checks every dispatch against the agent’s declared tier. It warns by default; AIGENT_MODEL_GUARD=enforce blocks.
• Teams. There’s written guidance on when to run a standing team, a one-shot fan-out or a solo agent. That’s guidance only, no code.
• Codex adapter. Routes one bounded, mechanical task type to codex exec and writes a diff for review. It never commits or pushes.

Running four seats is operating experience, not a measured result. The README does not claim aigent-os grew the newsletter or cut operator effort.

Install

Full install (needs Claude Code, Node.js 18+ with npm, and bash; on Windows use Git Bash):

git clone https://github.com/wrg32786/aigent-os.git && cd aigent-os && bash install.sh --dry-run

The dry run prints the planned changes and writes nothing. When it looks right, run bash install.sh, open a new terminal and run aigent.

Just want the manual capsule commands? Install the context-capsule plugin inside Claude Code:

/plugin marketplace add wrg32786/aigent-os
/plugin install context-capsule@aigent-os

Limits

• The public evidence is 11 of 11 refresh cycles on one Windows seat, in one live run during v1.1.0-rc.1 testing. That is single-seat evidence, not a reliability claim, and not a claim that this beats compaction.
• The managed runner can still miss a permission dialog and answer it (#104). Until the fix lands, don’t run it unattended on a seat that can hit permission prompts.
• A seat launched from inside another Claude Code session can lose its transcript (#98). Launch seats from a plain terminal.
• Scrolling the seat window during a refresh sends stray input (#99).
• macOS has the least live mileage with the managed runner.

Full list: the Known limits section of the v1.1.0 release notes.

• The fleet-band and fleet-delivery mods (status line, message delivery into an idle session) are an opt-in pilot on master, not in a release, and delivery needs a message room and job source that aren’t in the public repo. fleet-delivery can type into a seat during a refresh (#105).

FAQ

Can I run Claude Code as a multi-agent system with separate memory?

Yes. Each aigent-os seat has its own memory vault and capsules. The maintainers run four seats on one business.

Does it work with Codex or other models?

Claude Code is the host. A Codex adapter routes one bounded task class to codex exec, review gated and never auto merged. Other CLIs are on the roadmap, not shipped.

Is it free?

aigent-os is free and MIT licensed. You pay for Claude Code as usual, and each refresh uses some tokens to write and read the capsule.

More

• Repo and full README: https://github.com/wrg32786/aigent-os
• All question pages
• Releases