0.6.2

Self-Evolving RAG for AI Agents

Agents don't just read memory — they write it.

中文 | English

Core Concept

User Input → Retrieve KB → Generate
               ↑
            Read-only
        (Human-maintained)

User Input → Retrieve Memory → Generate
               ↑↓
           Read + Write
    Agent autonomously evolves

Key Differences:

How It Works

In Claude Code: "Remember, I prefer TypeScript"
         ↓
    Agent automatically calls:
    • append_daily_log() → Record to daily log
    • update_preference() → Update preferences
    • extract_knowledge() → Extract knowledge graph
         ↓
In Antigravity: "What are my coding preferences?"
         ↓
    AI: "You prefer TypeScript"

Teach once, remember forever. Share across apps, never forget.

Getting Started

Prerequisites

1. Python 3.10+
2. Ripgrep (rg): REQUIRED for full-text search. (Windows: choco install ripgrep, macOS: brew install ripgrep)
3. SQLite: Handled automatically by Python.

Configuration (v0.6.0)

Configuration is managed via Environment Variables.

1. mcp.json Structure

{
  "mcpServers": {
    "adaptive-agent-mcp": {
      "command": "uvx",
      "args": ["adaptive-agent-mcp"],
      "env": {
        "ADAPTIVE_EMBEDDING_BASE_URL": "https://api.xxx.cn/v1",
        "ADAPTIVE_EMBEDDING_API_KEY": "sk-your-xxx-key",
        "ADAPTIVE_EMBEDDING_MODEL": "Qwen/Qwen2.5-Coder-7B-Instruct",
        "ADAPTIVE_RERANK_BASE_URL": "https://api.xxx.cn/v1",
        "ADAPTIVE_RERANK_API_KEY": "sk-your-xxx-key",
        "ADAPTIVE_RERANK_MODEL": "BAAI/bge-reranker-v2-m3"
      }
    }
  }
}

Local Models:

• Ollama: Set ADAPTIVE_EMBEDDING_PROVIDER to ollama.
• LM Studio/vLLM: Set ADAPTIVE_EMBEDDING_PROVIDER to openai_compatible.
• Base URL: Set to your local endpoint (e.g., http://localhost:11434/v1 or http://localhost:1234/v1).
• API Key: Any string.

2. Environment Variables

All variables are prefixed with ADAPTIVE_.

Default storage path: ~/.adaptive-agent/memory. All apps share the same memory.

Enhance Agent Memory Behavior (Optional)

If your AI doesn't actively read/write memory, add this to your system prompt or user rules:

## Memory System Instructions

- At the start of each conversation, call `initialize_session` to load user preferences.
- When user says "remember", "save", or expresses preferences, call `update_preference` or `append_daily_log`.
- After completing tasks, briefly record progress using `append_daily_log`.
- When user asks about past conversations, use `query_memory_headers` or `search_memory_content`.

Features

Available Tools (14 tools)

Session & Retrieval

Memory & Knowledge

Knowledge Graph

Storage Structure

~/.adaptive-agent/memory/
├── MEMORY.md                          # User preferences (scope-based)
├── knowledge/
│   └── areas/
│       ├── general/items.json         # Global knowledge
│       ├── chat/items.json            # Chat-scope knowledge
│       ├── coding/items.json          # Coding-scope knowledge
│       ├── writing/items.json         # Writing-scope knowledge
│       └── projects/{name}/items.json # Project-specific knowledge
├── .index/
│   ├── vectors.db                     # SQLite + sqlite-vec + FTS5
│   └── index.json                     # Indexer metadata
├── .graph/
│   └── knowledge.json                 # NetworkX graph
├── .locks/                            # File lock directory
└── memory/
    └── 2026/
        └── 02_february/
            └── week_07/
                └── 2026-02-10.md      # Daily logs

Data Safety

• Isolated storage: Data stored in ~/.adaptive-agent/memory, independent of uvx installation
• Concurrent safety: filelock prevents data corruption from multiple clients
• Human-readable: All data in Markdown/JSON format, easy to backup and version control

License

MIT License - See LICENSE for details.