Global Agent Memory
Local-first Markdown MCP memory with human review, Obsidian vault, and owner dashboard.
Global Agent Memory is an open-source local-first memory system (PyPI `global-memory-mcp` 0.1.6, MIT) that keeps Markdown as the canonical vault, exposes a frozen MCP v1 contract to Claude Code, Codex, and other MCP clients, and lets owners approve candidates, protect sensitive memories, and administer projects from a local dashboard—with SQLite FTS5 keyword search and optional Ollama + sqlite-vec hybrid ranking.
Official resources
Selection advice
Quick comparison
Global Agent Memory centers a human-reviewed Markdown vault shared across MCP coding agents. EverOS also treats Markdown as source of truth, while Mem0 embeds memory primarily through application SDKs.
| Global Agent Memory | EverOS | Mem0 | |
|---|---|---|---|
| Best for | Shared, reviewable project memory across Claude Code, Codex, and other MCP clients | Portable Markdown-native memory OS across local agent workflows | Embedding user memory inside a product you ship |
| Storage model | Canonical Markdown vault + rebuildable SQLite/FTS indexes; optional Ollama vectors | Markdown source of truth with SQLite and LanceDB indexes | SDK-managed memory with self-hosted and managed options |
| Governance | Candidates require owner approval; Standard/Protected/Sealed visibility levels | Local-first runtime with reflection and knowledge-wiki surfaces | Application-controlled APIs rather than an owner review dashboard |
| Tradeoff | Strong local control and review workflow; requires installing and running a local daemon | Broad coding-assistant portability; less emphasis on candidate approval queues | Fast product embedding; less file-native Obsidian-style ownership |
Best for
- local-first shared agent memory
- human-reviewed durable knowledge
- Obsidian-compatible Markdown vaults
- MCP coding agents on macOS or Linux
Not ideal for
- teams that want a fully managed memory SaaS
- Windows-only workflows (installer targets macOS and Linux)
- products that need an embeddable hosted memory API for end users
Core concepts
Minimal implementation shape
Install with `uv tool install global-memory-mcp`, run `global-memory setup`, then ask a connected agent to load project context or propose a candidate via MCP tools such as `memory_context` and `memory_remember`.