# superlocalmemory [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/varun369/SuperLocalMemoryV2  
**GitHub Stars:** 255  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/superlocalmemory

## Description
Local-first AI memory with knowledge graphs and hybrid search. 17+ AI tools via MCP. Free.

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

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

## Documentation

## What superlocalmemory MCP server does

The superlocalmemory MCP server provides persistent memory for agents that otherwise lose context when a session ends. It stores facts such as decisions, rules, corrections, and project information, then retrieves relevant entries through MCP tools and the command-line interface.

The system combines knowledge-graph relationships with hybrid search. Retrieval uses multiple channels, result fusion, and a learned ranker rather than relying only on vector similarity. Memories also retain event time and learning time, allowing queries about what was true at a particular point or what the system knew before a date.

Access is organized around workspaces, roles, and personal, shared, or global scopes. The material describes default-deny cross-profile recall, retention rules, HMAC-verifiable erasure receipts, and a hash-chained audit log. Agent identity is recorded for writes, and trust scoring is used to address potentially harmful memory contributions.

## How it works

The superlocalmemory MCP server can run in Mode A, where core remember and recall operations make no model-provider call. The software runs locally on Windows, Linux, and macOS, and does not require Docker, a graph database, or an API key. Data leaving the machine is described as an explicit choice rather than a requirement of core retrieval.

When an agent recalls information, an answer-check stage evaluates whether the returned memories actually address the question. Laya performs that check locally on a Mac, while Jev runs online on Windows, Linux, and macOS. If the retrieved material does not answer the question, recall can report an abstention instead of presenting an unrelated memory as an answer.

The system can learn from retrieval outcomes. A Thompson-sampling bandit adjusts channel weights, and a LightGBM ranker uses reported outcomes. New rankers undergo a shadow comparison before promotion and can be rolled back when the documented NDCG@10 threshold is missed.

## Setup and configuration

Install the package with npm globally or with pipx. The documented npm route requires Node 18 or later and Python 3.12 or later; the pip route uses an activated Python virtual environment. After installation, run `slm setup` and select a mode. The command `slm connect <client>` connects the installation to a supported MCP client, such as Cursor or Claude Code.

A repository-based installation is also documented through platform-specific shell scripts for macOS, Linux, and Windows. The application supports all three operating systems.

## Tools and capabilities

The superlocalmemory MCP server exposes 17 or more AI tools through MCP. The documented core operations include:

- Saving memories with categories such as decisions, rules, facts, and corrections.
- Recalling memories with ranked results and optional kind filters.
- Replacing an existing fact and reviewing or rolling back that correction.
- Querying memories by validity time or knowledge date.
- Applying access scopes, retention rules, erasure workflows, and audit records.
- Sharing messages, locks, and state between parallel agent sessions through SLM-Mesh.
- Using exact caching and reversible compression through MCP tools or a skill.
- Repeating bounded tasks until an independent test, schema, or linter gate passes, with each attempt recorded.

## Limitations and notes

Search scores indicate ranking, not probability. Recall may show `Incomplete search` while the embedding model is loading. Answer checking depends on the selected judge: Laya is local to Mac, while Jev is online.

_Full upstream README: https://allmcps.com/mcp/superlocalmemory/readme_