convez.ai
MCP ServerProdutividadeIntermediário
AWDDude

engRam — Servidor MCP de Memória Semântica Local

AWDDude/engRam

Memória semântica persistente em um único binário Go, sem serviços externos

Ver no GitHub
InicialPluginScore · de 0 a 100

Um servidor MCP compilado em Go que implementa memória de longo prazo com busca semântica sem dependências externas. Usa embeddings locais via hugot e persiste em banco vetorial embarcado (chromem-go), oferecendo 5 ferramentas (store, search, list, update, delete) para armazenar e recuperar informações contextualizadas.

Instalação

$npx convez-ai AWDDude/engRam

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
v3.0.0
Issues abertas
5
Contribuidores
1

Casos de uso

  • Manter contexto pessoal e preferências do usuário entre sessões
  • Armazenar fatos, tarefas e ações para recuperação semântica
  • Implementar memória de longo prazo em agentes Claude com busca inteligente

Veredito

Solução madura e bem-engineered para memória local sem overhead de serviços externos. Ideal para quem quer estender Claude com contexto persistente; requer conforto com binários Go e configuração JSON, mas abstrai complexidade de embeddings e vetorizações.

README

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

engRam

A long-term semantic memory MCP server — single statically-linked Go binary with no external services required.

Name: engram (a memory trace in neuroscience) + RAM — stores memories, fast.

Features

  • Single binary — no Python, no Docker, no runtime dependencies
  • Local embeddings via hugot + GoMLX (all-MiniLM-L6-v2, Apache 2.0, downloaded once on first run)
  • Persistent vector search via chromem-go (embedded, file-based)
  • 5 MCP tools — store, search, list, update, delete
  • Semantic search with optional type filtering
  • XDG-compliant data and config paths on all platforms
  • Model migration — switch embedding models without losing memories

Installation

Homebrew (macOS / Linux)

brew install AWDDude/tap/engram

Build from source

git clone https://github.com/AWDDude/engRam.git
cd engRam
make build     # produces ./engram (static binary, CGO_ENABLED=0)

Then copy the binary somewhere on your PATH, e.g. ~/.local/bin/.

Internet access is required on first run to download the embedding model from Hugging Face (~90MB, one-time only).

MCP configuration

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "engram": {
      "command": "engram"
    }
  }
}

If you built from source or installed to a custom path, use the full path to the binary instead.

Configuration

engRam uses XDG-style directories by default on all platforms:

PurposeDefault path
Data (db, models)~/.local/share/engram/
Config~/.config/engram/config.json

XDG_DATA_HOME and XDG_CONFIG_HOME are honored on all platforms. Override the config path entirely with ENGRAM_CONFIG_PATH.

If the config file does not exist, engram creates it with defaults on first run. All fields are required — engram will log any missing fields and exit if the file is incomplete.

{
  "model": {
    "path": "/path/to/models",
    "embedding_model": "KnightsAnalytics/all-MiniLM-L6-v2"
  },
  "db": {
    "path": "/path/to/db"
  }
}

Warning: changing model.embedding_model requires a migration — engram will refuse to start and tell you to run engram migrate.

Switching embedding models

If you change model.embedding_model in your config, engram will detect the mismatch on startup and return an error:

engram: embedding model changed from "KnightsAnalytics/all-MiniLM-L6-v2" to "your/new-model"
— run 'engram migrate' to re-embed all memories

Run the migration command to re-embed all memories with the new model:

engram migrate

Migration is atomic — the new collection is fully built before the old one is removed. If it fails partway through, your existing memories are untouched.

Tools

ToolRequiredOptional
storecontent, typetags
searchquerymin_score (default 0.5)
list—type_filter, tag_filter, limit (default 20)
deletememory_id—
updatememory_id, content—

Memory types: preference, task, fact, action

min_score is a cosine similarity threshold (0–1). Results below it are excluded. Omit to use the default of 0.5.

delete and update return an error if memory_id does not exist.

Examples

store(content="prefers dark mode", type="preference", tags=["ui"])
search(query="UI preferences")
search(query="UI preferences", min_score=0.7)
list(type_filter="fact", tag_filter="kubernetes")
update(memory_id="<id>", content="updated content")
delete(memory_id="<id>")

Development

make test    # go test -v -race ./...
make build   # CGO_ENABLED=0 static binary
make clean   # remove binary

Tests use an in-memory store with a deterministic mock embedding function — no model download required.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados