# gingugu/gingugu [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/gingugu/gingugu  
**GitHub Stars:** 4  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/gingugu-gingugu

## Description
Persistent memory for AI coding assistants. Local SQLite, no cloud. 16 MCP tools: store, recall, search, relate, consolidate, export, and credential vault (OS keychain). Typed memories with confidence lifecycle (verified/inferred/stale/deprecated), namespaces, knowledge graph, and hybrid BM25 + semantic search via fastembed ONNX. Works with Cursor, Windsurf, Claude, and any MCP client. pip install gingugu

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "gingugu": {
    "command": "uvx",
    "args": ["gingugu"]
  }
}
```

## Documentation

## What gingugu/gingugu MCP server does

The gingugu/gingugu MCP server provides persistent memory for AI coding assistants through MCP. Memory remains available across sessions, repositories, and projects instead of being limited to one conversation or client. Data is stored locally in a SQLite file, with no cloud service, API key, or telemetry requirement.

Memories can be organized by namespace and type, assigned confidence states, and connected in a typed knowledge graph. Supported lifecycle states include verified, inferred, stale, and deprecated. The system can also surface review hints for time-sensitive memories without silently changing their contents.

## How it works

The server combines SQLite FTS5 BM25 search with local semantic embeddings and merges the results using Reciprocal Rank Fusion. The default embedding option uses fastembed with an ONNX model and can work offline after its initial model download. An Ollama backend is also available when an existing Ollama process should provide embeddings. Semantic search can be disabled to use BM25-only retrieval.

Returned results can be ranked using relevance, freshness, access patterns, and confidence. Session-start behavior can load relevant namespaces and inject memory guidance through a hook. A stop hook checks whether a session involving real work recorded anything, helping make memory capture part of the workflow rather than an occasional manual action.

## Setup and configuration

The project requires Python 3.11 or newer. The README documents installation with `pip install gingugu`. Running `gingugu init` connects a repository to the workflow and installs the session hooks described above.

Storage is local and does not require account credentials. The optional embedding settings include `MEMORY_EMBEDDINGS_BACKEND=ollama` for Ollama or `MEMORY_EMBEDDINGS_ENABLED=false` to turn semantic search off. The default fastembed model is downloaded to the local fastembed cache and then operates offline.

## Tools and capabilities

The gingugu/gingugu MCP server exposes operations for storing, recalling, searching, relating, consolidating, and exporting memories. Its capabilities include:

- Namespace-scoped memory for repositories, projects, and shared patterns
- Typed relationships such as supersedes, contradicts, caused_by, and parent/child links
- Duplicate and related-memory scans, followed by merge, summary, or deduplication actions
- Session-start context loading with optional compact output
- Memory statistics, dormancy reports, review sweeps, and namespace summaries
- Secure service-bundle storage for API keys and tokens through the operating system keychain
- An interactive memory explorer for viewing the knowledge graph and dashboard

## Limitations and notes

The project describes itself as suitable for local personal workflows but still early for broad validation, large-scale databases, and long upgrade horizons. SQLite FTS5 is expected to handle millions of rows, but the project has not benchmarked memory stores at the 100,000-plus scale. The published retrieval measurements come from an internal labeled workload and should be treated as a regression baseline, not a cross-product benchmark.

The server is designed for local use and supports Cursor, Windsurf, Claude, and other MCP clients. The optional Ollama backend adds a dependency on a running Ollama process; fastembed avoids that requirement but downloads a model initially. The gingugu/gingugu MCP server does not require cloud credentials for its local operation.

_Full upstream README: https://allmcps.com/mcp/gingugu-gingugu/readme_