convez.ai
MCP ServerAutomaçãoAvançado
tangruiyuanbhe123456

Hermes Memory – Sistema de Memória Multicamadas para Agentes IA

tangruiyuanbhe123456/hermes-memory

Memória inteligente offline para agentes IA com autorreflexão e consolidação

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema de memória offline-first com 5 camadas (working, episódic, semantic, procedural, metacognitive) para agentes IA, incluindo busca FTS5, servidor MCP, dashboard web, consolidação de sonhos e motor de autorreflexão. Zero dependências externas, criptografia AES-256-GCM e suporte a múltiplos agentes (Claude, Codex, Hermes).

Instalação

$npx convez-ai tangruiyuanbhe123456/hermes-memory

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 4 semanas
Versão
v2.3.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Agentes IA que precisam aprender com histórico de ações
  • Sistemas multi-agente com memória compartilhada
  • Análise de estratégias e pitfalls em workflows de automação
  • Busca semântica em conhecimento acumulado sem APIs externas

Veredito

Projeto robusto e diferenciado que resolve um problema real: memória persistente e inteligente para agentes IA sem dependências cloud. A camada L5 de autorreflexão é inovadora. Requer razoável curva de aprendizado, mas entrega valor sólido para quem constrói sistemas multi-agente complexos.

README

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

Hermes Memory

Five-layer memory system for AI agents. Zero external dependencies. FTS5 full-text search. MCP protocol. Web dashboard. Dream consolidation. AES-256-GCM encryption.

🚀 npx skills init → then npx skills install memory-system


Why Hermes Memory?

Most AI agent memory systems fall into two camps:

  1. Cloud-dependent (Mem0, Supermemory) — need API keys, vector DBs, cloud infra
  2. Too simple — just save/recall, no search, no dedup, no strategy

Hermes Memory is the only offline-first, zero-dependency memory system with a self-reflection (L5) engine — it doesn't just remember, it learns which strategies work and which don't.

FeatureHermes MemoryMem0LettaSupermemoryKeyMemory
Offline-only, zero external deps✅❌✅❌✅
FTS5 Chinese+English search✅❌❌❌❌
Self-reflection engine (L5)✅❌❌❌❌
Dream consolidation (dedup+merge)✅❌❌❌✅
Write-before-write auto backup✅❌❌❌✅
Web UI dashboard✅❌❌✅✅
MCP server (any agent)✅❌❌❌✅
Multi-agent (Codex+Claude+Hermes)✅❌✅❌✅
Platform health tracking✅❌❌❌❌
Pitfall knowledge base✅❌❌❌❌

Architecture

hermes-memory/
├── engine/
│   ├── memory-db.cjs      ← Core: SQLite FTS5 engine + backup + encryption
│   ├── dashboard.cjs      ← Web UI (http://127.0.0.1:3210)
│   ├── mcp-server.cjs     ← MCP protocol server
│   ├── dream.cjs          ← Dream consolidation engine
│   └── agent-config.cjs   ← Agent config generator
├── memory/                ← 5-layer memory data
│   ├── L1_working/        ← Session state
│   ├── L2_episodic/       ← Operation timeline
│   ├── L3_semantic/       ← Long-term knowledge
│   ├── L4_procedural/     ← Reusable skills
│   └── L5_metacognitive/  ← Self-reflection, strategy evaluation
├── agent-configs/         ← Ready-to-use configs for Claude/Codex/Hermes
├── codex-memory/          ← Codex-compatible memory (same 5 layers)
├── package.json
└── memory.db              ← SQLite FTS5 index (auto-generated)

The Five Layers

LayerNamePurpose
L1WorkingCurrent session state, scratchpad
L2EpisodicAction log with timestamps (JSONL + FTS5)
L3SemanticLong-term facts, patterns, knowledge
L4ProceduralReusable workflows, skills, recipes
L5MetacognitiveStrategy analysis: what worked, what didn't, what to try next

Quick Start

Install

# Via skills.sh (recommended for AI agents)
npx skills init
npx skills install memory-system

# Via npm (standalone)
npm install hermes-memory
npx hermes-memory-dashboard

Start the Dashboard

node engine/dashboard.cjs
# → http://127.0.0.1:3210

Connect Your AI Agent

Hermes Agent — install as a skill (see above) Claude Code — cp agent-configs/CLAUDE.md /path/to/project/CLAUDE.md Codex — the MCP server auto-detects Codex sessions Any MCP-compatible agent — point to the MCP server:

{
  "mcpServers": {
    "hermes-memory": {
      "command": "node",
      "args": ["path/to/engine/mcp-server.cjs"]
    }
  }
}

Run Dream Consolidation

# Preview (dry-run)
node engine/dream.cjs --dry-run

# Execute
node engine/dream.cjs

Health Check

node -e "const e=require('./engine/memory-db.cjs'); console.log(JSON.stringify(e.dbHealth(),null,2))"

Key Features

🔍 FTS5 Chinese + English Search

Unlike vector DB-based systems, Hermes Memory uses SQLite FTS5 with a CJK-padding technique that lets you search mixed Chinese+English queries like "小红书发布pending" with zero external dependencies.

🧠 L5 Self-Reflection Engine

The metacognitive layer tracks which strategies succeed and fail over time. It automatically recommends optimal approaches based on historical patterns. No other memory system has this.

💭 Dream Consolidation

Automatically deduplicates, merges, and archives old memories. Keeps your knowledge base lean and relevant without manual cleanup.

🔐 AES-256-GCM Secrets

Store API keys and credentials encrypted at rest. No .env file needed.

💾 Auto-Backup

Every write to L3/L4/L5 is automatically backed up with 7-day retention. Transaction-safe restore.


Comparison with Competitors

CapabilityHermes MemoryMem0SupermemoryLettaKeyMemory
StarsNew58K26K23KNew
Offline, zero deps✅❌❌✅✅
CJK FTS5 search✅❌❌❌❌
L5 self-reflection✅❌❌❌❌
Dream consolidation✅❌❌❌✅
Web dashboard✅❌✅❌✅
MCP protocol✅❌❌❌✅
Write backup✅❌❌❌✅
Platform health✅❌❌❌❌
Pitfall KB✅❌❌❌❌
npm install✅❌✅✅❌

License

MIT

Ver o original no GitHub

Bom para

Relacionados