Agentic Stack: A portable brain for your AI coding assistants

Keep your AI memory across multiple coding harnesses

The Agentic Stack repository introduces a groundbreaking, portable .agent/ folder designed to standardize memory, skills, and protocols across multiple AI coding harnesses like Claude Code, Cursor, and Windsurf. By centralizing the AI’s learning process and utilizing host-agent review tools, developers can prevent knowledge loss when switching between different development toolkits. This tool represents a strategic step toward harness-agnostic AI workflows, ensuring long-term compounding of machine-learned development contexts.

Points clés

Authored by AV1DLIVE based on “The Agentic Stack” concept and initially coded using Minimax-M2.7 in the Claude Code harness
Supports seamless integration across 8 distinct AI harnesses, including Claude Code, Cursor, Windsurf, OpenCode, OpenClaw, Hermes, Pi Coding Agent, and standalone Python
Introduces a highly portable .agent/ folder structure containing four separate memory layers: working, episodic, semantic, and personal
Features an auto_dream.py nightly staging cycle that mechanically clusters recurring patterns into candidate lessons for the host agent
Replaces automated reasoning with a mandatory host-agent review protocol, requiring developers to manually accept or reject AI-generated candidate lessons with a provided rationale
Version 0.7.0 introduces host-agent tools like learn.py for one-shot lesson teaching and a colorful dashboard show.py to manage the AI’s brain state
Includes five foundational seed skills upon installation, such as skillforge for creating new skills and git-proxy to manage Git operations safely
Offers an opt-in full-text search capability through the FTS5 beta feature to rapidly query all markdown and JSONL memory documents

À retenir

If you are tired of your AI assistant suffering from digital amnesia every time you switch platforms, it might be time to pack its bags into an Agentic Stack. While setting up an intricate system of episodic memories and nightly “dreams” sounds a lot like you are raising a very needy digital toddler, it definitely beats teaching your code editor how to write your commit messages for the hundredth time. Just remember, the AI still needs you to manually rubber-stamp its daily lessons, so you probably shouldn’t plan your early retirement to a tropical island just yet.

Sources