Coding agents that rewrite their own memory while you sleep

What it does

This TypeScript harness runs agents that persist state, skills, and conversation history across sessions. Instead of treating context as ephemeral, it stores memory in a git-backed filesystem called MemFS, so an agent retains its identity even when you swap underlying models or move between devices. Interaction happens through a CLI, desktop app, browser, or messaging integrations like Slack and Telegram.

The interesting bit

Built by the same researchers behind MemGPT and sleep-time compute, the project treats memory as the primary architecture, not a bolt-on. Agents can trigger periodic “dreaming” to rewrite their own prompts, audit memory health with /doctor, and invoke themselves or other agents as subagents to handle tasks asynchronously.

Key highlights

• MemFS tracks all context in git and can sync to a private GitHub repository, making agent memory version-controlled.
• A self-improvement loop lets agents learn new skills and update their own system prompts without manual intervention.
• Built-in subagent types—recall, history-analyzer, forked workers—allow agents to delegate work and search across conversations.
• Runs fully offline via an embedded local server, or connects to Constellation for cross-machine state sync and remote execution environments.
• Hooks, permission modes, crons, and cross-agent message search provide automation and guardrails.

Verdict

Try it if you want a coding partner that accumulates institutional knowledge rather than resetting every morning. Avoid it if you prefer stateless, one-shot LLM interactions or are uncomfortable with an agent that maintains its own evolving git history.