AgentSkel: An Open-Source Framework for Building Persistent Memory and Team-Shared Standards for AI Programming Agents

AgentSkel is an innovative open-source framework designed to solve the 'session amnesia' problem of AI programming agents. By storing memory files via Git branches, standardizing workflows, and establishing a cross-tool compatible rule system, AgentSkel allows every AI agent in a team to share project knowledge, follow unified standards, and maintain context continuity across multiple sessions. It addresses the fundamental limitation of AI programming assistants (like Claude Code, GitHub Copilot, Cursor) — session statelessness — which leads to repeated work, inconsistent behaviors, and lack of shared architectural understanding in team environments.

AI programming assistants are revolutionizing software development but suffer from session statelessness: each new session requires re-scanning codebases, forgetting prior conventions, repeating errors, and ignoring past architectural decisions. In team settings, this worsens: agents behave inconsistently, propose conflicting solutions, and lack unified understanding of shared architectures (e.g., iOS/Android agents diverging on business logic).

Without AgentSkel, common issues include:

Issue	Symptom
Session Amnesia	Agents re-scan 40 files per session and repeatedly ask already answered questions
Codebase Scale Limitation	Large codebases exceed context windows, forcing agents to guess
Lack of Institutional Knowledge	Agents are unaware of edge cases and reasons behind design decisions
Standard Drift	5 developers +5 AI tools =5 different agent behaviors
Cross-Platform Divergence	iOS and Android agents silently diverge on shared business logic
Repeated Mistakes	No agent learns from last week’s errors

AgentSkel solves these issues via three core mechanisms:

1. Persistent Knowledge Base: Uses Git isolated branch (ai-memory) as .memory/ worktree to store project knowledge in pure Markdown files.
2. Always-Enforced Rules: Hard principles like pre-coding planning, pre-release validation, and non-negotiable security.
3. Structured Workflows: 17 standardized processes covering feature development, bug fixes, code reviews, etc.

Key architectural design:

• Isolated Branch Worktree: Main branch holds app code, ai-memory branch stores memory files (mounted to .memory/), ensuring version control, non-intrusiveness, no vendor lock-in, and cross-tool compatibility.
• Memory File System: Includes MAP.md (code structure), SYMBOLS.md (class/function index), RESUME.md (session state), RULES.md (core rules), CONVENTIONS.md (coding styles), etc.

Structured workflows include:

• Core: Feature development (plan → branch → TDD → PR), bug fix (reproduce → test → root cause → fix), brainstorming.
• QA: Code review, test engineering, tech debt repair.
• Ops: Release, hotfix, dependency check.
• Knowledge: Codebase mapping, subagent dispatch, navigation.

AgentSkel supports multiple AI tools:

Tool	Integration Method
Claude Code	Plugin + session-start hook
Cursor	Plugin + native rules
GitHub Copilot	.github/copilot-instructions.md
Windsurf	Native rules
Codex CLI	Natively reads AGENTS.md
Gemini / Antigravity	Extension + GEMINI.md

Compared to system prompts:

Feature	System Prompt	AgentSkel
Sharing	Personal level	Team level (shared via Git)
Persistence	Resets on session end	Permanently stored in Git
Cross-Tool	Incompatible	All tools read the same files
Knowledge Accumulation	Linear	Compound growth (no repeated errors)
Architectural Decision Propagation	Manual notification	Auto-synced

Installation:

• Claude Code: /plugin install agentskel
• Gemini CLI: gemini extensions install https://github.com/ahmadulhoq/agentskel
• Manual: git clone https://github.com/ahmadulhoq/agentskel.git → ./scripts/install-agent.sh

Project Initialization: Open project directory and say: "Set up AgentSkel for this project." The agent creates memory files, copies rules/workflows, and opens a PR. After merging, say "Map this codebase" to build the shared knowledge base.

Version Management: Agents detect version gaps at session start. Use: "Sync this project with the latest skeleton." to apply updates and open PRs.

Implications:

• From personal tool to team infrastructure.
• From session memory to institutional knowledge.
• From tool lock-in to open standards.
• From linear improvement to compound growth.

Limitations:

1. Learning curve for team adoption of conventions.
2. Maintenance cost for updating memory files.
3. Partial tool support for specific features.
4. Need for Git advanced feature familiarity.

AgentSkel provides an elegant solution to AI programming agents' amnesia problem. It uses Git branches for persistent memory, standardized workflows for team consistency, and cross-tool compatibility to avoid vendor lock-in. As AI-assisted programming becomes widespread, AgentSkel will be a critical component for scaling AI collaboration in teams, turning individual efficiency tools into shared infrastructure.