Memory Forge
БесплатноНе проверенForge persistent memories for your AI agent. 9 MCP tools + 5 auto-engines. Powered by Shelby decentralized cloud.
Описание
Forge persistent memories for your AI agent. 9 MCP tools + 5 auto-engines. Powered by Shelby decentralized cloud.
README
Persistent memory engine for AI agents.
One command. Zero config. 8 MCP tools + 5 auto-engines.
Powered by Shelby Protocol — decentralized hot storage on Aptos.
The Problem
Your AI agent has amnesia.
Every time you close Claude Code or Cursor, your AI agent forgets everything. You told it you prefer React 19 with TypeScript. You taught it your project's auth flow. You spent 20 minutes debugging a config issue together. Close the terminal — gone.
There are 25+ MCP memory servers on the market. Most store memories in local files. Lose your machine, lose your agent's brain.
What Makes MemoryForge Different
1. Local-First, Cloud-Optional
Free tier runs entirely on your machine. Zero network requests. Upgrade to Pro for Shelby decentralized cloud sync across devices.
2. Actually Automatic
5 background engines handle naming, deduplication, priority scoring, decay archiving, and context injection. You never run memory_* commands — your agent calls them automatically.
3. PreCompact Auto-Capture
Before context compaction, MemoryForge prompts your agent to save key learnings. Even if you force-close the terminal, your memories survive.
4. Zero External Dependencies
Embedding model (23MB Transformers.js) runs in-process. No API keys needed for Free tier. Works offline.
Quick Start
npx memory-forge setup
Auto-installs hooks, imports your existing rules, preloads the embedding model. Works with Claude Code, Cursor, Windsurf, VS Code.
Pro (Shelby Cloud Sync)
SHELBY_API_KEY="your-key" memory-forge pro
Cross-device memory sync via Shelby decentralized storage (currently on Shelbynet testnet).
How to Use
Install (one-time)
npx memory-forge setup
Daily Use — Nothing Required
- Session start → context auto-injected
- During conversation → agent calls
memory_storeandmemory_searchautomatically - PreCompact → auto-capture instruction prompts agent to save key learnings
- Session end → autoPriority recalc + autoDecay archiving
Talk to Your Agent Naturally
"Remember: our project uses pnpm, not npm"
"Search for memories about authentication"
"List all my memories"
"Share this memory with the frontend team"
"Export all memories as JSON backup"
Your agent calls the right MCP tool automatically.
8 MCP Tools
| Tool | Description |
|---|---|
memory_store |
Store with auto-embedding, auto-naming, auto-dedup |
memory_search |
Semantic search — vector first, keyword fallback |
memory_recall |
Exact retrieval by memory ID |
memory_list |
Browse with category/tag filters and pagination |
memory_forget |
Delete memory (local file + in-memory cache) |
memory_context |
Load session context — top-N by access + priority |
memory_export |
Export as JSON or Markdown |
memory_share |
Package for teammate import |
5 Auto-Engines
| Engine | Trigger | Effect |
|---|---|---|
| Auto-Name | Every memory_store |
Extracts descriptive names from content |
| Auto-Merge | Every memory_store |
Merges >80% similar memories |
| Auto-Priority | Session Stop | Scores 1–10 from frequency × recency × age |
| Auto-Decay | Session Stop | Ebbinghaus curve: 7d→0.8, 30d→0.5, 90d→archive |
| Context Injection | SessionStart hook | Loads top-5 relevant memories each session |
Pricing
| Free | Pro | |
|---|---|---|
| Price | $0 | Testnet (free) |
| Storage | Local Markdown | Shelby decentralized cloud |
| Memory limit | Unlimited | Unlimited |
| Cross-device sync | Manual export | Auto-sync |
| Tools | All 8 | All 8 |
| Auto-engines | All 5 | All 5 |
Architecture
┌──────────────────────────────────────────────┐
│ Agent (Claude Code / Cursor / Windsurf) │
└──────────────┬───────────────────────────────┘
│ MCP (JSON-RPC over stdio)
▼
┌──────────────────────────────────────────────┐
│ MemoryForge MCP Server │
│ ┌──────────┐ ┌───────────┐ ┌────────────┐ │
│ │ 8 Tools │ │ 5 Engines │ │ CLI (setup │ │
│ │ store │ │ auto-name │ │ pro hook) │ │
│ │ search │ │ auto-merge│ │ │ │
│ │ recall │ │ priority │ │ memory- │ │
│ │ list │ │ decay │ │ forge ... │ │
│ │ forget │ │ context │ │ │ │
│ │ context │ └───────────┘ └────────────┘ │
│ │ export │ │
│ │ share │ TypeScript, Node 18+, ESM │
│ └──────────┘ │
│ │
│ ┌──────────────────────────────────────┐ │
│ │ Embedding Engine │ │
│ │ Transformers.js (23MB in-process) │ │
│ │ all-MiniLM-L6-v2, 384-dim vectors │ │
│ │ Fallback: keyword matching │ │
│ └──────────────────────────────────────┘ │
└──────────┬───────────────┬───────────────────┘
│ │
▼ ▼
┌──────────────────┐ ┌─────────────────────────┐
│ Free Storage │ │ Pro Storage │
│ ~/.memory-forge │ │ Shelby Protocol │
│ memories/*.md │ │ @shelby-protocol/sdk │
│ (Markdown) │ │ Shelbynet testnet │
└──────────────────┘ └─────────────────────────┘
Security
| Data | Location | Notes |
|---|---|---|
| Memory content | ~/.memory-forge/memories/ (Free) or Shelby blob (Pro) |
User-managed |
| API keys | Environment variable only | Never written to disk |
| Private keys | ~/.memory-forge/pro.json (Pro) |
User-local |
| Payment info | Not applicable | No payment system yet |
Free tier: zero network requests (except one-time 23MB model download). Pro tier: HTTPS transport, on-chain Aptos transactions.
Tech Stack
| Layer | Technology |
|---|---|
| Protocol | @modelcontextprotocol/sdk (MCP stdio) |
| Embeddings | @huggingface/transformers (Xenova/all-MiniLM-L6-v2, 23MB) |
| Validation | zod |
| Free Storage | Local Markdown (~/.memory-forge/memories/*.md) |
| Pro Storage | @shelby-protocol/sdk + @aptos-labs/ts-sdk |
| Runtime | Node.js 18+, TypeScript, ESM |
Zero external SaaS. Zero API keys required for Free. Embedding model runs locally.
Competitive Landscape
| MemoryForge | Mem0 | Zep | LangMem | Pinecone | |
|---|---|---|---|---|---|
| Storage | Local + Decentralized | Cloud | Cloud | Local | Cloud |
| Protocol | MCP | REST | REST | SDK | gRPC |
| Cross-device | ✅ (Pro) | ✅ | ✅ | ❌ | ❌ |
| In-process embedding | ✅ (23MB) | ❌ | ❌ | ❌ | ❌ |
| Auto-engines | 5 | 2 | 1 | 0 | 0 |
| Zero API keys | ✅ (Free) | ❌ | ❌ | ✅ | ❌ |
MemoryForge is the only MCP memory server combining local-first privacy with optional decentralized cloud storage.
Roadmap
v0.1.x ✅ Published — 8 tools + 5 auto-engines + local storage
v0.2.0 ✅ All-English docs, PreCompact auto-capture
v0.2.x ⬜ MCP directory listing (mcp.so, smithery.ai, glama.ai)
v0.3.0 ⬜ SettleGrid payment integration
v1.0.0 ⬜ Mainnet launch, web dashboard
Documentation
| Document | Content |
|---|---|
| TECHNICAL.md | API reference, data model, architecture, security |
| TUTORIAL.md | Install guide, daily use, Pro setup, troubleshooting |
| SPEC.md | Product specification and roadmap |
| ARCHITECTURE.md | System architecture and data flow |
Links
License
MIT © shelby-protocol
Установить Memory Forge в Claude Desktop, Claude Code, Cursor
unyly install memory-forgeСтавит в Claude Desktop, Claude Code, Cursor и VS Code — сам разбирается с npx, uvx и сборкой из исходников.
Впервые? Поставь CLI: curl -fsSL https://unyly.org/install | sh
Или настроить вручную
Выполни в терминале:
claude mcp add memory-forge -- npx -y memory-forgeFAQ
Memory Forge MCP бесплатный?
Да, Memory Forge MCP бесплатный — установка в пару кликов через Unyly без оплаты.
Нужен ли API-ключ для Memory Forge?
Нет, Memory Forge работает без API-ключей и переменных окружения.
Memory Forge — hosted или self-hosted?
Self-hosted: сервер запускается локально на твоей машине командой из раздела установки.
Как установить Memory Forge в Claude Desktop, Claude Code или Cursor?
Открой Memory Forge на unyly.org, выбери вкладку своего клиента (Claude Desktop, Claude Code, Cursor) и нажми Install — конфиг сгенерируется автоматически, без правки JSON.
Похожие MCP
GitHub
PRs, issues, code search, CI status
автор: GitHubFilesystem
Secure file operations with configurable access controls.
Memory
Knowledge graph-based persistent memory system.
Template MCP Server
A CLI tool to create a new Model Context Protocol server project with TypeScript support, dual transport options, and an extensible structure
автор: mcpdotdirectCompare Memory Forge with
Не уверен что выбрать?
Найди свой стек за 60 секунд
Автор?
Embed-бейдж для README
Похожее
Все в категории development
