Mimir MCP - Git-Backed AI Memory System

Mimir MCP - Git-Backed AI Memory System

Research Date: 2026-02-04 Source URL: https://github.com/tejzpr/mimir-mcp GitHub Repository: https://github.com/tejzpr/mimir-mcp Docker Hub: https://hub.docker.com/r/tejzpr/mimir-mcp Version at Research: Latest (no tagged releases) License: MPL-2.0 (Mozilla Public License 2.0)

Overview

Mimir is a Model Context Protocol (MCP) server that provides persistent, git-versioned memory storage for LLM applications. It combines Git's version control with graph-based memory associations, enabling LLMs to maintain long-term memory with full audit trails and relationship tracking.

Core Value Proposition: Transform ephemeral AI conversations into persistent, interconnected knowledge bases with complete version history and relationship graphs.

Problem Addressed

Key Statistics (as of 2026-02-04)

Note: Very new project (1 day old at research time). Statistics expected to change rapidly.

Key Features

Memory Management

• 7 Human-Aligned MCP Tools: Tools express intent rather than implementation
  • mimir_recall - "What do I know about X?" (semantic search)
  • mimir_remember - "Store this for later" (create/update)
  • mimir_history - "When did I learn about X?" (temporal queries)
  • mimir_connect - "These are related" (link memories)
  • mimir_forget - "No longer relevant" (soft delete/archive)
  • mimir_restore - "Bring back that archived memory" (undelete)
  • mimir_sync - "Sync with remote" (manual git sync)

Storage Architecture

• Git-Backed Primary Storage: Every memory change is a commit
• SQL Index: SQLite (development) or PostgreSQL (production)
• Dual Storage: Git repository (primary) + SQL database (index)
• Auto-Sync: Hourly synchronization to GitHub with PAT authentication
• Encryption: 32-character encryption key for sensitive data

Memory Format

Memories stored as Markdown files with YAML frontmatter:

---
id: project-alpha-kickoff-2024-01-15
title: "Project Alpha Kickoff Meeting"
tags: [project, meeting]
created: 2024-01-15T10:30:00Z
updated: 2024-01-15T14:00:00Z
associations:
  - target: contact-john-doe
    type: person
    strength: 1.0
---

# Project Alpha Kickoff Meeting
[Content...]

Graph Associations

• 7 Relationship Types: related, references, follows, supersedes, part_of, person, project
• Typed Connections: Link memories with semantic relationships
• N-Hop Queries: Traverse memory associations across multiple levels
• Supersedes Chain: Automatically marks old memories as outdated when replaced

Authentication

• Local Mode: Development use with system user (whoami)
• SAML 2.0: Production authentication with DUO/Okta support
• Multi-User Mode: ACCESSING_USER environment variable for user isolation
• Token Management: Configurable token TTL (default 24 hours)

Deployment Options

• Native Binary: Build from source with make build
• Go Run: go run github.com/tejzpr/mimir-mcp/cmd/server@latest
• Docker Container: docker run tejzpr/mimir-mcp
• stdio Mode: Default for MCP client integration
• HTTP Mode: Web interface with authentication UI

Technical Architecture

┌─────────────────────────────────────────────────────────────┐
│                      MCP Client                              │
│              (Cursor, Claude Desktop, etc.)                  │
└─────────────────────────────────────────────────────────────┘
                              │
                    stdio / HTTP transport
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                     Mimir MCP Server                         │
│                                                              │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐  │
│  │ mimir_recall│  │mimir_remember│  │  mimir_connect     │  │
│  │ mimir_history│ │mimir_forget │  │  mimir_sync        │  │
│  │ mimir_restore│ │             │  │                     │  │
│  └─────────────┘  └─────────────┘  └─────────────────────┘  │
│                                                              │
│  ┌─────────────────────────────────────────────────────┐    │
│  │              Memory Service Layer                    │    │
│  │   - Search (topic, exact, list_all, path)           │    │
│  │   - CRUD operations with slugs                      │    │
│  │   - Association management                          │    │
│  │   - History and temporal queries                    │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘
                              │
          ┌───────────────────┼───────────────────┐
          ▼                                       ▼
┌─────────────────────────┐           ┌─────────────────────────┐
│   SQL Database Index    │           │   Git Repository        │
│                         │           │                         │
│  - SQLite (dev)         │           │  ~/.mimir/store/        │
│  - PostgreSQL (prod)    │           │    mimir-{username}/    │
│  - Full-text search     │           │      2024/01/           │
│  - Relationship graph   │           │      tags/              │
│                         │           │      archive/           │
└─────────────────────────┘           └─────────────────────────┘
                                                  │
                                        Auto-sync (hourly)
                                                  │
                                                  ▼
                                      ┌─────────────────────────┐
                                      │    GitHub Remote        │
                                      │    (with PAT auth)      │
                                      └─────────────────────────┘

Directory Structure

~/.mimir/
├── configs/
│   └── config.json              # User configuration
├── db/
│   └── mimir.db                 # SQLite database (index)
└── store/
    └── mimir-{username}/        # User's git repository
        ├── 2024/
        │   └── 01/              # Date-organized memories
        ├── tags/
        │   └── meetings/        # Tag-based organization
        └── archive/             # Soft-deleted memories

Installation and Usage

Build from Source

make setup    # Run initial setup
make deps     # Install dependencies
make build    # Build the binary

MCP Client Configuration

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "mimir": {
      "command": "/path/to/bin/mimir",
      "env": {
        "ENCRYPTION_KEY": "your-32-char-encryption-key-here"
      }
    }
  }
}

Docker Mode:

{
  "mcpServers": {
    "mimir": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "/Users/yourname/.mimir:/home/mimir/.mimir",
        "-e", "ENCRYPTION_KEY=your-32-char-encryption-key-here",
        "tejzpr/mimir-mcp"
      ]
    }
  }
}

Tool Usage Examples

Recall (Search):

{
  "topic": "authentication approach",
  "limit": 10
}

Remember (Store):

{
  "title": "Project Alpha Kickoff",
  "content": "# Meeting Notes\n\nDiscussed project timeline...",
  "tags": ["project", "meeting"],
  "replaces": "old-meeting-notes"
}

Connect (Link):

{
  "from": "project-alpha",
  "to": "contact-john-doe",
  "relationship": "person"
}

History (Temporal):

{
  "slug": "project-alpha",
  "show_changes": true,
  "since": "7d"
}

Relevance to Claude Code Development

Direct Applications

1. Cross-Session Memory: Enable Claude Code to remember decisions, solutions, and context across conversations
2. Project Knowledge Base: Build persistent, searchable knowledge graphs for complex projects
3. Decision Audit Trail: Track why decisions were made with full git history
4. Team Knowledge Sharing: Multi-user mode enables shared memory across team members

Patterns Worth Adopting

1. Human-Aligned Tool Design: Tools express intent ("What do I know about X?") rather than implementation ("SELECT FROM memories WHERE...")
2. Supersedes Relationships: Clean handling of outdated information without deletion
3. Git-Based Persistence: Leverage existing version control for AI memory durability
4. Dual Storage Architecture: Index for speed, git for durability and audit
5. Proactive Memory Protocol: "Recall before answering, store after solving"

Integration Opportunities

1. Claude Code Plugin: Could extend existing context-management approaches in this repository
2. Skill Memory: Skills could persist learned patterns and user preferences
3. Session Continuity: Hook into session start/end for automatic memory sync
4. Research Curation: Use for curating and linking research findings (like this directory)

Comparison to Existing Solutions

Prompt Integration

The repository includes sample prompts for AI assistants:

Key Behaviors (from SAMPLE_MCP_PROMPT.md)

1. Check first - Use mimir_recall before answering questions
2. Store valuable info - Decisions, solutions, context, action items
3. Supersede, don't duplicate - Use replaces param when updating
4. Connect while storing - Use connections param to link memories

Do Not Store

• Credentials/secrets
• Anything user doesn't want stored

References

1. GitHub Repository: https://github.com/tejzpr/mimir-mcp (accessed 2026-02-04)
2. Docker Hub Image: https://hub.docker.com/r/tejzpr/mimir-mcp (accessed 2026-02-04)
3. README Documentation: https://raw.githubusercontent.com/tejzpr/mimir-mcp/main/README.md (accessed 2026-02-04)
4. Sample MCP Prompt: https://raw.githubusercontent.com/tejzpr/mimir-mcp/main/SAMPLE_MCP_PROMPT.md (accessed 2026-02-04)
5. Sample Cursor Rule: https://raw.githubusercontent.com/tejzpr/mimir-mcp/main/SAMPLE_CURSOR_RULE.md (accessed 2026-02-04)
6. GitHub API Repository Data: https://api.github.com/repos/tejzpr/mimir-mcp (accessed 2026-02-04)

Freshness Tracking

Change Detection Indicators:

• Monitor GitHub releases for version tags
• Check Docker Hub for image updates
• Watch for star growth indicating adoption
• Review for production readiness indicators
• Check for additional MCP tool additions