Tool Pick | mcp-memory — Long-Term Agent Memory Using Google's OKF Standard

🌏 中文版

Tool Info

What Problem Does It Solve

Have you ever spent a whole session getting an Agent to understand your project architecture, your coding style preferences, and the progress on that bug you're tracking — only to start a new conversation the next day and find it's a blank slate again? Or maybe you're using Claude Code, Cursor, and Codex simultaneously, each with its own siloed "memory" (Claude's CLAUDE.md is plain text, Cursor has its own rules file), with no way to share a single long-term memory across tools.

mcp-memory's approach is to define "memory" as a standardized format rather than a proprietary feature of any single tool. It adopts Google Cloud Platform's Open Knowledge Format (OKF v0.2) — each memory entry is a Markdown file with YAML frontmatter (type, key, namespace, tags, status, generated, verified, etc.), stored in a memory/ folder in your project. Humans can open and read the files directly, and Git can diff them. A separate SQLite FTS5 index (.mcp_memory/memories.db) enables millisecond full-text search by keyword or tag, instead of scanning all Markdown files on every query.

Good fit for: developers maintaining a long-lived project who want the Agent to remember architecture decisions and personal preferences; people using multiple MCP-compatible Agents (Claude Desktop, Cursor, Antigravity, Windsurf, Codex) who want them to share a single memory store instead of each maintaining its own; and teams that care about memory being Git-versionable, human-readable, and not locked into any vendor's format.

Getting Started

Installation

# Requires Python 3
git clone https://github.com/fellowgeek/mcp-memory
cd mcp-memory
python3 setup.py

setup.py is an interactive wizard that auto-detects and registers the server in your installed Claude Desktop, Cursor, Antigravity, Windsurf, or Codex config files. You can also manually add the run.sh path to your MCP client config. The entire process requires no account registration, no API key — all data stays local.

Basic Usage

The Agent gets 5 MCP tools:

• memory_store — create or update a memory entry in OKF v0.2 format
• memory_retrieve — fetch a specific memory by key + namespace
• memory_search — search memories by keyword, tag, or namespace
• memory_get_last — read the last working progress checkpoint
• memory_update_last — update the current progress checkpoint

A stored memory looks like this (memory/user/preferences/coding_style.md):

---
type: Agent Memory
title: Coding Style
key: user/preferences/coding_style
namespace: default
tags:
  - preferences
  - style
status: stable
generated:
  by: mcp-memory/0.2.0
  at: '2026-08-12T19:23:35Z'
created_at: '2026-08-12T19:23:35Z'
updated_at: '2026-08-12T19:23:35Z'
---

User prefers functional programming style with explicit type annotations.

Advanced Usage

Use namespace for context isolation — separating "user preferences" from "project architecture" memories. You can also use environment variables to customize storage locations (e.g., sharing a single memory store across multiple projects, or moving the SQLite index to a different disk):

export MCP_MEMORY_PROJECT_ROOT=/path/to/project
export MCP_MEMORY_DIR=memory                  # OKF Markdown storage folder
export MCP_MEMORY_DB_PATH=.mcp_memory/memories.db  # SQLite index location

Comparison with Existing Tools

Caveats

• This project was published just 4 days ago: most of the 175 stars came from a single Hacker News spike, and long-term maintenance remains to be seen. Read through db.py and memory_server.py source code before adopting in production.
• OKF itself is still early-stage (v0.2): this is a recently published standard from Google Cloud Platform, and the schema may change. mcp-memory currently tracks SPEC.md / OKF_RULES.md, but if the upstream spec revises, this server will need to follow suit to avoid producing incompatible memory files.
• No memory_delete tool: you can only create, update, search, and read checkpoints. Removing a memory entry requires manually deleting the Markdown file under memory/ and rebuilding the SQLite index — cleaning up erroneous memories is inconvenient for now.

Takeaway

I used to think "Agent memory" was a problem everyone solved their own way — Mem0 uses a vector database, Claude uses a 200-line-capped Markdown file, LangMem is tied to the LangChain ecosystem. But mcp-memory reminded me that people are already pushing for an open standard for the memory format itself (OKF), trying to let different Agent toolchains share a single memory store instead of each inventing their own format — much like how every browser had its own markup before the web standards movement. Standardization is still very early, but the direction itself is worth watching.

References

• fellowgeek/mcp-memory — GitHub
• Open Knowledge Format v0.2 SPEC — GoogleCloudPlatform/knowledge-catalog
• Show HN: MCP Memory – Fast Agent Memory Using Google's OKF and SQLite FTS5 — Hacker News