Skillget

claude-engram

mcp
Guvenlik Denetimi
Basarisiz
Health Gecti
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 14 GitHub stars
Code Basarisiz
  • eval() — Dynamic code execution via eval() in mini_claude/handlers.py
Permissions Gecti
  • Permissions — No dangerous permissions requested
Purpose

This tool provides persistent memory and context tracking for AI coding assistants. It automatically logs mistakes, decisions, and session states across multiple projects using background hooks and local semantic search.

Security Assessment

Overall Risk: Medium

The server is designed to be local-only and does not request inherently dangerous system permissions. However, the rule-based scan flagged a critical code vulnerability: the use of dynamic code execution via `eval()` in one of its handler files. The use of `eval()` is a high-risk pattern that can lead to arbitrary code execution if an attacker finds a way to manipulate the input passed to it. While the tool does not appear to make external network requests or contain hardcoded secrets, the `eval()` vulnerability requires caution. Additionally, it relies on a local Ollama instance for semantic search, which developers should ensure is securely configured.

Quality Assessment

The project is in active development, with its last push occurring today. It is properly licensed under the permissive and standard MIT license. Community trust is currently low but growing, represented by 14 GitHub stars. The codebase is open-source, allowing developers to easily inspect the implementation and verify exactly how the `eval()` function is being utilized.

Verdict

Use with caution due to the presence of dynamic code execution (`eval()`) in the codebase.
SUMMARY

Persistent memory for Claude Code. Auto-tracks mistakes, decisions, and context via hooks. Tiered storage, semantic scoring, multi-project workspaces. Local-only.

README.md

Mini Claude

Persistent memory and self-awareness for Claude Code. Tracks mistakes, decisions, and context across sessions automatically via hooks. Includes loop detection, scope guards, tiered memory with archiving, semantic intent scoring, and a local LLM for code analysis.

Features

  • Auto-tracks mistakes from any failed tool. Warns before editing the same file again.
  • Auto-captures decisions from user prompts ("let's use X instead of Y") via semantic + regex scoring.
  • Detects edit loops when the same file is edited 3+ times without progress.
  • Survives compaction by auto-saving checkpoints and re-injecting rules/mistakes after.
  • Saves session handoffs so the next session picks up where you left off.
  • Archives old memories to cold storage instead of deleting. Searchable, restorable.
  • Scores and injects context before every edit, surfacing the 3 most relevant memories.
  • Scopes memory per project in multi-project workspaces. Workspace-level rules cascade down.

Most of this works automatically via Claude Code hooks. No tool invocations needed.

Install

# 1. Ollama (for semantic search / code analysis)
ollama pull gemma3:12b

# 2. Clone and install
git clone https://github.com/20alexl/mini_claude.git
cd mini_claude
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows

pip install -e .                # Base
pip install -e ".[semantic]"    # + AllMiniLM decision capture (recommended)

# 3. Configure hooks + MCP server
python install.py

Per-Project Setup

python install.py --setup /path/to/your/project

Or copy .mcp.json and CLAUDE.md to your project root manually.

Quick Usage

After install, Mini Claude works automatically. You'll see hook output like:

Mini Claude session started (startup)
Rules (2):
  [a1b2c3] Always use strict TypeScript
Past mistakes (1):
  [d4e5f6] Broke the auth middleware by removing the session check
Edit tracked: server.py (edit #2)
FAIL Test tracked
Auto-logged: Import error: Module 'flask' not found

For manual operations:

memory(operation="remember", content="...", project_path="/path")
memory(operation="add_rule", content="Always do X", project_path="/path")
work(operation="log_decision", decision="...", reason="...")
scout_search(query="how does auth work", directory="/path")

Configuration

export MINI_CLAUDE_MODEL="gemma3:27b"           # Ollama model (default: gemma3:12b)
export MINI_CLAUDE_OLLAMA_URL="http://host:11434" # Remote Ollama
export MINI_CLAUDE_KEEP_ALIVE="5m"               # Keep model loaded
export MINI_CLAUDE_ARCHIVE_DAYS=30               # Archive threshold (default: 14)
export MINI_CLAUDE_SCORER_TIMEOUT=3600           # Scorer idle timeout (default: 1800)

Troubleshooting

Problem Fix
MCP server not connecting Check ollama list, restart Claude Code
Hooks not firing Run python install.py to reinstall
Memories not showing memory(operation="archive_status", project_path="/path") to check counts
ModuleNotFoundError Activate the venv first

Documentation

Read the Library Book for the complete guide: design principles, internals, full usage guide, advanced features, gotchas, and complete tool reference.

License

MIT

Yorumlar (0)

Sonuc bulunamadi