Your AI assistant’s memory, packed for travel across IDEs

What it does

agentic-stack drops a .agent/ directory into a project and treats it as a portable brain: memory layers, skill definitions, and protocol files that wire into Claude Code, Cursor, Windsurf, Copilot CLI, Gemini CLI, and others via manifest-driven adapters. Instead of each tool maintaining its own siloed context, the project keeps one shared state that follows you across harnesses. It also includes local-first data skills that generate cross-agent dashboards, cost estimates, and daily reports, plus a flywheel skill that packages approved, redacted runs into trace records and training-ready JSONL without external APIs.

The interesting bit

The project essentially decouples your agent’s identity from the IDE or CLI running it. Upgrades are surgical—new infrastructure files land without touching personal memory, skill directories, or adapter settings—so your agent’s autobiography doesn’t get rewritten by a package update. The optional bridge to an external git-backed brain server even lets context survive across different projects, not just different tools.

Key highlights

• Adapter system covers 12+ harnesses, from mainstream IDEs to DIY Python loops, managed through a shared Python backend with concurrent-write protection.
• Local-first data skills generate unified dashboards, cron timelines, token/cost estimates, and daily reports across all wired agents.
• Data-flywheel skill exports approved, redacted traces into eval cases and training-ready JSONL without training models or sending telemetry.
• Safe upgrade path previews skeleton changes and leaves user memory, CLAUDE.md, and existing skills untouched.
• Optional external Brain integration adds cross-project, git-backed long-term memory via an MCP bridge.

Caveats

• Several capabilities are explicitly beta and off by default: tldraw visual canvas, Mission Control web dashboard, and FTS memory search.
• The external Brain memory layer requires installing a separate Rust CLI; it is not bundled.
• Adapter removal is immediate with “no quarantine, no undo,” per the README.

Verdict

Worth a look if you are juggling multiple AI coding tools and tired of re-teaching preferences and project context. Probably overkill if you are monogamous with a single agent that already remembers everything you need.