nexus-memory
A zero-dependency, file-based persistent memory system for AI agents with tiered memory, Ebbinghaus decay, and keyword retrieval.
README
Nexus Memory System
Zero-dependency, file-based persistent memory for AI agents.
Nexus is a tiered memory system with Ebbinghaus decay, keyword retrieval, token-efficient context assembly, and full MCP + REST API support. Drop-in replacement for Hindsight that saves 92% on memory token costs.
MCP Server → stdio (Claude Code, Cline, Windsurf)
REST API → HTTP (Hermes agents, custom integrations)
CLI → bash (nexus.sh — search, stats, decay, consolidate)
Features
| Feature | Description |
|---|---|
| Zero dependencies | No database, no vector store, no embeddings API. Pure Python stdlib. |
| MCP native | 5 tools (search, stats, save, touch, decay) + resource access |
| REST API | Hindsight v1 compatible. Drop-in replace HINDSIGHT_API_URL. |
| Ebbinghaus decay | Automatic forgetting curve. Memories expire on schedule. |
| Token economics | 92% cost reduction vs Hindsight. Built-in token tracking. |
| Pointer-based RAG | Kronos-style 300-token pointers for budgeted context assembly. |
| File-based | Plain markdown files. Readable, editable, git-versionable. |
| Bilingual | Full Chinese + English support. |
| Cross-agent sharing | Share memories across Hermes agents or any MCP client. |
Quick Start
# 1. Start the MCP server (for Claude Code / Cline / Windsurf)
python nexus_mcp.py
# 2. Start the REST API (for Hermes agents / HTTP clients)
python nexus_rest.py --port 9177
# 3. Use the CLI
python nexus_engine.py retrieve "what do I know about X"
python nexus_engine.py stats
python nexus_engine.py decay
Claude Code Integration
Add to your claude.json:
{
"mcpServers": {
"nexus-memory": {
"command": "python",
"args": ["path/to/nexus_mcp.py"]
}
}
}
Hermes Agent Integration
Replace Hindsight with Nexus:
export HINDSIGHT_API_URL=http://localhost:9177
No code changes needed. Nexus speaks the Hindsight v1 protocol.
Architecture
┌─────────────────────────────────────────────────────┐
│ Nexus System │
│ │
│ ┌──────────────┐ ┌──────────┐ ┌───────────────┐ │
│ │ nexus_mcp.py │ │nexus_rest│ │nexus_engine.py│ │
│ │ (MCP stdio) │ │(HTTP API)│ │ (Core logic) │ │
│ └──────┬───────┘ └────┬─────┘ └───────┬───────┘ │
│ └───────────────┼─────────────────┘ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ memory/ (files) │ │
│ │ ├ episodic/ │ │
│ │ ├ semantic/ │ │
│ │ ├ procedural/ │ │
│ │ ├ reflections/ │ │
│ │ ├ working/ │ │
│ │ ├ core/ │ │
│ │ └ archive/ │ │
│ └──────────────────┘ │
└─────────────────────────────────────────────────────┘
Memory Tiers
| Tier | Decay | Purpose |
|---|---|---|
| Working | 7 days | In-session context |
| Episodic | 30 days | Past experiences |
| Semantic | 90 days | Facts, preferences |
| Procedural | 180 days | Workflows, skills |
| Reflections | 60 days | Meta-cognition |
| Core | Never | Identity, rules |
Token Economics
| Metric | Hindsight | Nexus | Savings |
|---|---|---|---|
| Per recall | 500 tokens | 30 tokens | 94% |
| Per retain | 300 tokens | 50 tokens | 83% |
| 5 agents/day | 440,000 tokens | 36,000 tokens | 92% |
| Monthly cost | $39.60 | $3.24 | $36.36 |
Benchmark: 1192.9x efficiency ratio (1 token spent → 1192 saved vs Hindsight).
Pricing
Free Solo Team Enterprise
───── ────── ────── ──────────
Price $0 $4.99/mo $14.99/mo $49.99/mo
Memories 50 500 5,000 50,000
MCP ✓ ✓ ✓ ✓
REST API ✓ ✓ ✓ ✓
CLI ✓ ✓ ✓ ✓
Pointers - ✓ ✓ ✓
Token 7 days 30 days 90 days 365 days
tracking
Cross- - - ✓ ✓
agent
Priority - - - ✓
support
All tiers include Ebbinghaus decay, keyword retrieval, and file-based transparency.
Roadmap
- [x] MCP server (tools + resources)
- [x] REST API (Hindsight v1 compatible)
- [x] Keyword retrieval + scoring
- [x] Token economics tracking
- [x] Ebbinghaus decay
- [x] Memory consolidation
- [ ] x402 micropayments
- [ ] SSE transport for MCP
- [ ] Cloud sync
- [ ] Knowledge graph
Why Not Hindsight?
Hindsight is powerful but expensive: it calls LLMs for every recall/retain, uses PostgreSQL + pgvector, and requires a running daemon. Nexus achieves comparable retrieval quality at 8% of the token cost — no LLM calls, no database, no daemon. Just files and algorithms.
Why Not Mem0/Letta/Memoria?
Those are excellent systems, but they're architecturally heavy (vector DBs, embeddings, graph stores). Nexus is designed for the 80% use case: fast keyword retrieval with smart ranking. When you need semantic search, Nexus pointers bridge the gap at zero marginal cost.
No database. No API keys. No Docker. Just python nexus_mcp.py.
Built with ❤️ for the Hermes agent ecosystem.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。