convez.ai
MCP ServerBusca & PesquisaIntermediário
gelutjari

AIKD – Memória Instantânea para Agentes de IA

gelutjari/aikd

Memória híbrida BM25+vetorial para agentes de código com latência de 0,21ms

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP Server ultrarrápido que oferece aos agentes de IA acesso instantâneo à memória semântica e BM25 do seu codebase em Rust. Funciona 100% localmente, sem envio de código para a nuvem, indexando 10 mil chunks em 0,21ms com setup em apenas 30 segundos.

Instalação

$npx convez-ai gelutjari/aikd

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 5 meses
Versão
—
Issues abertas
14
Contribuidores
1

Casos de uso

  • Consultas semânticas em grandes codebases durante debugging e refatoração
  • Manutenção de contexto persistente entre sessões do agente de IA
  • Busca local de padrões arquiteturais sem exposição de código na nuvem

Veredito

Solução elegante e performática para o problema real de amnesia contextual em agentes de IA. A implementação em Rust com indexação híbrida (BM25+vetorial) e execução 100% local resolve limitações críticas de ferramentas existentes, sendo ideal para equipes que priorizam privacidade e latência baixa.

README

A tradução para o português deste README ainda está na fila.

🧠 AIKD

The Ultra-Fast Local Memory Layer for AI Coding Agents

Crates.io GitHub Release Build Status License: MIT Rust Downloads

Give Claude, Cursor, and Cline instant memory of your entire codebase. Written in Rust. Zero cloud dependency. Search 10,000 chunks in 0.21ms.

Install • Quick Start • Demo • Docs


🤔 Why AIKD?

The Problem: AI coding agents are powerful but forgetful. Every new conversation starts from zero — no memory of your codebase, your patterns, your architecture.

Pain PointWithout AIKDWith AIKD
😵 Context AmnesiaAgent asks "what does this file do?" every sessionInstant recall of entire codebase
🐌 Slow SearchPython-based RAG takes 500ms+ per queryRust-powered BM25+Vector in 0.21ms
☁️ Cloud DependencyYour code sent to external servers100% local, zero data leaves your machine
🔧 Complex SetupInstall Python, pip, CUDA, models...Single binary, one command install

How AIKD Compares

FeaturegrepLlamaIndexai-devkitAIKD
SpeedFastSlowMedium⚡ 0.21ms
Semantic Search❌✅✅✅ Hybrid
Local Only✅❌❌✅ 100%
MCP Native❌❌❌✅ Built-in
Memory UsageLowHigh (Python)High📊 27% RAM
Setup Time0s~10min~5min⏱️ 30s

✨ Key Features

🔍 Search Engine

  • Hybrid Search: BM25 (Tantivy) + Vector (ONNX all-MiniLM-L6-v2, 384d)
  • Reciprocal Rank Fusion: Combines keyword + semantic results intelligently
  • Incremental Indexing: Blake3 hashing for zero-redundancy scans
  • Resource Adaptive: Auto-detects CPU/GPU and adjusts batch sizes

🤖 AI Integration

  • MCP Protocol: Native support for Model Context Protocol
  • 7 Built-in Tools: scan, query, embed, stats, remember, recall, status
  • Auto-Registration: Works with Claude Code, Cursor, Cline, Continue, Windsurf, MiMoCode
  • Session Memory: Persistent conversation context across sessions

🛠️ Developer Experience

  • Single Binary: No Python, no Node, no dependencies
  • File Watcher: Auto-reindex on file changes
  • REST API: HTTP endpoint on port 9090 for custom integrations
  • CLI First: Full control from terminal

🎬 See It In Action

AIKD Demo

$ aikd init
✅ AIKD initialized for project

$ aikd scan
📊 Indexed 847 files, 12,453 chunks in 1.2s

$ aikd query "authentication flow"
1. src/auth/login.rs (Lines 45-89, Score: 0.923)
   Implements JWT-based authentication with refresh tokens...

2. src/middleware/auth.rs (Lines 12-34, Score: 0.871)
   Validates Bearer tokens on protected routes...

3. docs/architecture/auth.md (Lines 1-25, Score: 0.845)
   Authentication flow diagram and security considerations...

📦 Installation

Recommended (One-Line)

Linux / macOS:

curl -sSf https://raw.githubusercontent.com/gelutjari/aikd/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/gelutjari/aikd/main/install.ps1 | iex

Alternatives

cargo install aikd
brew install gelutjari/tap/aikd
git clone https://github.com/gelutjari/aikd.git
cd aikd
cargo build --release
cp target/release/aikd ~/.local/bin/
docker run -v $(pwd):/workspace -p 9090:9090 ghcr.io/gelutjari/aikd:latest

⚡ Quick Start (3 minutes)

Step 1: Initialize

cd your-project
aikd init

Creates ~/.aikd/config.yaml with smart defaults based on your project type (Rust, Node, Python, Go). It detects .git, Cargo.toml, package.json, etc. and sets appropriate file filters.

Step 2: Scan Your Codebase

aikd scan
  1. Walks directory tree (skipping node_modules, .git, target, etc.)
  2. Chunks files by semantic boundaries (functions, headings)
  3. Generates BM25 index (Tantivy) and vector embeddings (ONNX)
  4. Stores everything in local SQLite database

Step 3: Search

# Keyword search (fast)
aikd query "error handling"

# Semantic search (smart)
aikd query "how does login work" --hybrid

# Get JSON output for scripting
aikd query "database connection" --json

Step 4: Connect to Your AI Agent

AIKD auto-registers via MCP. Just restart your AI agent after running aikd scan.

The MCP config is at ~/.aikd/mcp.json:

{
  "mcpServers": {
    "aikd": {
      "command": "aikd",
      "args": ["serve"]
    }
  }
}
# Start the daemon
aikd daemon

# Query via HTTP
curl "http://localhost:9090/api/query?q=authentication&limit=5"

🎛️ Usage Modes

                    ┌─────────────────┐
                    │   How to use    │
                    │     AIKD?       │
                    └────────┬────────┘
                             │
            ┌────────────────┼────────────────┐
            ▼                ▼                ▼
     ┌──────────┐     ┌──────────┐     ┌──────────┐
     │   CLI    │     │   MCP    │     │ REST API │
     │  Mode    │     │  Server  │     │  Daemon  │
     └──────────┘     └──────────┘     └──────────┘
     Best for:        Best for:        Best for:
     • Quick search   • AI agents      • Web apps
     • Scripts        • Claude Code    • Custom UIs
     • CI/CD          • Cursor         • Team shared

CLI Mode (default):

aikd query "rust error handling" --limit 5
aikd stats
aikd status

MCP Server (for AI agents):

aikd serve  # Starts stdio MCP server

REST Daemon (background service):

aikd daemon              # Start in background
aikd daemon foreground   # Start in foreground
aikd daemon-stop         # Stop daemon

📊 Benchmark Results

AIKD Benchmarks

Hardware: AMD EPYC 7B13, 7.8GB RAM, NVMe SSD

OperationTimeThroughput
Index 1,000 files144ms6,934 files/s
BM25 Search0.21ms4,762 queries/s
Hybrid Search0.35ms2,857 queries/s
Concurrent (500)28ms17,669 queries/s
Embedding (batch)12ms/batch~800 chunks/s

Resource Usage:

CPU: ████████░░░░░░░░░░░░ 22.9%
RAM: █████░░░░░░░░░░░░░░░ 27.4%

💡 Real-world context: A human eye blink takes ~300ms. In that time, AIKD can complete 857 hybrid searches across your entire codebase.


⚙️ Configuration

Default config (~/.aikd/config.yaml):

version: 2.0.0
scan:
  include_paths: ["."]
  exclude_paths: ["node_modules", ".git", "target"]
  include_extensions: ["rs", "ts", "py", "md", "json"]
  follow_symlinks: false
chunk:
  max_tokens: 1000
  min_tokens: 100
embedding:
  enabled: true
  model: all-MiniLM-L6-v2
  batch_size: auto
server:
  rest_port: 9090
  auth_token: null
  cors_origins: ["*"]
embedding:
  enabled: true
  batch_size: 8
  device: cpu
resource:
  mode: Low
embedding:
  enabled: true
  batch_size: 64
  device: gpu
resource:
  mode: Max
scan:
  include_extensions: ["md", "txt", "rst"]
  exclude_paths: [".git"]
filter:
  max_file_size: 524288  # 512KB
server:
  rest_port: 9090
  auth_token: "your-secret-token"
  cors_origins:
    - "https://your-team-app.com"

🏗️ Architecture

AIKD Architecture

┌─────────────────────────────────────────────────────────────┐
│                        AIKD v2.0.0                          │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────────┐  │
│  │  CLI    │  │  MCP    │  │  REST   │  │ File Watcher│  │
│  │ (clap)  │  │ (rmcp)  │  │ (axum)  │  │  (notify)   │  │
│  └────┬────┘  └────┬────┘  └────┬────┘  └──────┬──────┘  │
│       │            │            │               │          │
│       └────────────┴────────────┴───────────────┘          │
│                            │                                │
│  ┌─────────────────────────┴──────────────────────────┐    │
│  │                   Core Engine                       │    │
│  ├─────────────┬─────────────┬─────────────────────────┤    │
│  │  Scanner    │   Chunker   │      Session Manager    │    │
│  │ (walkdir)   │(pulldown-   │    (conversation DB)    │    │
│  │             │  cmark)     │                         │    │
│  └──────┬──────┴──────┬──────┴─────────────────────────┘    │
│         │             │                                      │
│  ┌──────┴──────┐ ┌────┴─────┐ ┌──────────────────────┐     │
│  │  Indexer    │ │ Embedder │ │     Storage           │     │
│  │  (Tantivy)  │ │ (ONNX)   │ │   (SQLite + WAL)     │     │
│  │  BM25 +     │ │ 384d     │ │   + r2d2 pool        │     │
│  │  HNSW ANN   │ │ MiniLM   │ │   + blake3 hashing   │     │
│  └─────────────┘ └──────────┘ └──────────────────────┘     │
│                                                             │
└─────────────────────────────────────────────────────────────┘

See docs/ARCHITECTURE.md for detailed breakdown.


🆘 Troubleshooting

Diagnostic Commands:

aikd status --json     # System info
aikd stats             # Index statistics
aikd daemon-pid        # Check if daemon running
  1. Run aikd scan first
  2. Check aikd stats to verify files indexed
  3. Try broader search terms
  4. Check include_extensions in config
aikd model download

This downloads ~90MB ONNX model to ~/.local/share/aikd/model/.

# Change port in config
# ~/.aikd/config.yaml
server:
  rest_port: 9091

Or stop existing process: aikd daemon-stop

Set resource mode to Low:

resource:
  mode: Low
  max_memory_mb: 512
  1. Check ~/.aikd/mcp.json exists
  2. Restart your AI agent
  3. Verify aikd serve works: echo '{"jsonrpc":"2.0","method":"initialize","id":1}' | aikd serve

Install Visual C++ Redistributable.


🤝 Community

Get Help:

Contributors:

ContributionWho
🏗️ Architecture & Core@gelutjari
🔒 Security AuditAI-assisted (MiMo)
⚡ Performance OptimizationCommunity
📖 DocumentationCommunity

See CONTRIBUTING.md for how to contribute.


📄 License

MIT License — see LICENSE for details.


⭐ Star this repo if AIKD helps your AI agents remember!

Star History Chart

Ver o original no GitHub

Bom para

Relacionados