# Semurg agent memory

Self-hosted memory for AI agents on your own hardware: a knowledge graph and vectors on one engine. Install it on Linux, then point your agent at it over MCP.

- MCP server · `type=mcp` · Claude Code, Codex, Cursor, GitHub Copilot, Windsurf (Devin Desktop), Cline, Zed
- Author: One-Semurg
- License: Apache-2.0 · policy=link-only · description only; the file stays with the author
- Source: https://github.com/One-Semurg/semurg-mcp
- Updated: 2026-10-24
- HTML: https://affpapa.org/en/nevibecode/mcp/semurg-agent-memory

## What it does

The node keeps notes and documents you hand to the agent. Each saved text is split into passages at sentence boundaries and linked to entities held in the graph, either already stored or sent alongside. The agent does the extracting itself by passing entities and relations in the same write call; the authors say that beats the node's own model running on CPU. Passages get vector embeddings when an embedding model is present on the machine.

Hardware first: x86_64 with AVX2 or newer, at least 2 CPU threads and 4 GB of RAM. Ubuntu 24.04 takes the installer; other x86_64 Linux distributions run in Docker. Nothing is published for Apple silicon, Windows or ARM, so from a Mac or a Windows box you reach a Linux node elsewhere. The data directory has to sit on a local ext4 or xfs volume, and the installer refuses to continue otherwise. Root is required, and a single node needs no licence key.

A scope is a namespace. Agents sharing a scope share facts; another scope stays private. In Claude the server is registered at user scope through `claude mcp add-json` and keeps only a reference to `SEMURG_API_TOKEN`. Other clients talk streamable HTTP to `POST /v1/mcp` with `Authorization: Bearer <token>` and an optional `X-Semurg-Scope`. The endpoint answers 401 without the token and 200 with it, so check that before trusting the setup.

## Install

### Claude Code (native platform)

**Install**
- The MCP registry has no launch command for it: see the repository README

## When to use it
- share one memory store across several agents
- attach entities and relations to text you save
- run vector search over notes without a hosted service
- keep each end user's memory apart inside one app
- confirm the MCP endpoint rejects requests without a token

## Example prompts
- Save this note and send the entities and relations as facts in the same `memory_remember` call
- Call `memory_status` for scope `default` and list the entity types it holds
- Search memory for the deploy notes and quote the passage the answer came from

## Similar
- [memory-systems](https://affpapa.org/en/nevibecode/skills/memory-systems.md) — skill · 393 installs
- [Synaplan](https://affpapa.org/en/nevibecode/mcp/synaplan.md) — MCP server · ★ 287
- [supermemory](https://affpapa.org/en/nevibecode/skills/supermemoryai-supermemory.md) — skill · 548 installs
- [shodh-memory](https://affpapa.org/en/nevibecode/mcp/shodh-memory.md) — MCP server · ★ 316
- [Local Memory](https://affpapa.org/en/nevibecode/mcp/local-memory.md) — MCP server · ★ 15
- [Pseudolife Memory](https://affpapa.org/en/nevibecode/mcp/pseudolife-memory.md) — MCP server · ★ 6

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md