MCP 服务器

adaptive-agent-mcp

A self-evolving RAG system that enables AI agents to autonomously read and write memory, continuously learning and adapting user preferences, daily logs, and knowledge graphs across applications.

README

Self-Evolving RAG for AI Agents

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

中文 | English

</div>

Core Concept

Traditional RAG

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

</td> <td width="50%">

Self-Evolving RAG

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

</td> </tr> </table>

Key Differences:

	Traditional RAG	Adaptive Agent MCP
Read	Retrieves pre-indexed documents	Dynamically accumulates at runtime
Write	Human-maintained knowledge base	Agent writes autonomously
Scope	Generic knowledge	User-specific memory
State	Static data	Continuously evolves

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

Python 3.10+
Ripgrep (rg): REQUIRED for full-text search. (Windows: choco install ripgrep, macOS: brew install ripgrep)
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_.

Variable	Description	Default
`ADAPTIVE_STORAGE_PATH`	Storage location	`~/.adaptive-agent/memory`
`ADAPTIVE_RIPGREP_PATH`	Path to `rg` executable	Auto-detect
`ADAPTIVE_EMBEDDING_PROVIDER`	Embedding provider (`openai_compatible`)	`openai_compatible`
`ADAPTIVE_EMBEDDING_BASE_URL`	API Endpoint	`None`
`ADAPTIVE_EMBEDDING_API_KEY`	API Key	`None`
`ADAPTIVE_EMBEDDING_MODEL`	Embedding Model	`Qwen/Qwen3-Embedding-8B`
`ADAPTIVE_RERANK_PROVIDER`	Rerank provider (`cohere_compatible`)	`cohere_compatible`
`ADAPTIVE_RERANK_BASE_URL`	API Endpoint	`None`
`ADAPTIVE_RERANK_API_KEY`	API Key	`None`
`ADAPTIVE_RERANK_MODEL`	Reranker Model	`Qwen/Qwen3-Reranker-8B`

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

Feature	Description	Version
Three-Layer Memory	MEMORY.md + Daily Logs + Knowledge Items	v0.1.0
Scope Isolation	`project:xxx`, `app:xxx`, `global`	v0.2.0
Concurrent Safety	Cross-process file locking + async locks	v0.3.0
Incremental Indexing	mtime-based smart updates	v0.3.0
Hybrid Search	Vector + FTS5 with RRF fusion	v0.6.0
Rerank Service	Cohere-compatible re-ranking for higher precision	v0.6.1
Area Partitioning	Scope-based knowledge routing	v0.6.0
Knowledge Graph	NetworkX-based entity relations	v0.5.0
Async Foundation	Non-blocking I/O throughout	v0.6.0

Available Tools (14 tools)

Session & Retrieval

Tool	Description
`initialize_session`	Initialize session with user profile and recent context
`query_memory_headers`	Index scan — browse memory file metadata
`read_memory_content`	Read complete memory file content
`search_memory_content`	Full-text search using ripgrep

Memory & Knowledge

Tool	Description
`update_preference`	Intelligently update user preferences
`append_daily_log`	Append content to daily log or knowledge items
`query_knowledge`	Hybrid search (Vector + FTS5 + RRF fusion) with browse fallback
`delete_knowledge`	Soft-delete knowledge items
`get_period_context`	Aggregate weekly/monthly logs for summaries
`archive_period`	Save period summaries

Knowledge Graph

Tool	Description
`extract_knowledge`	Extract entity relations from text
`add_knowledge_relation`	Manually add relations
`query_knowledge_graph`	Query entities, relations, or stats
`multi_hop_query`	Multi-hop reasoning queries

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