Claude Code / Opencode

Use mnemory with Claude Code or Opencode to remember project context, architecture decisions, coding preferences, and technical knowledge across sessions.

Setup

1. Start mnemory (if not already running):

   uvx mnemory
   

2. Add to your MCP configuration (Claude Code: ~/.claude/claude_code_config.json, Opencode: opencode.json):

Local setup (no auth):

{
  "mcpServers": {
    "mnemory": {
      "type": "streamable-http",
      "url": "http://localhost:8050/mcp",
      "headers": {
        "X-Agent-Id": "claude-code"
      }
    }
  }
}

Production (with auth):

{
  "mcpServers": {
    "mnemory": {
      "type": "streamable-http",
      "url": "https://mem.example.com/mcp",
      "headers": {
        "Authorization": "Bearer your-api-key",
        "X-Agent-Id": "claude-code"
      }
    }
  }
}

With MCP_API_KEYS configured on the server, user_id and agent_id are resolved automatically. No system prompt changes needed.

How It Works

Claude Code and Opencode natively support MCP server instructions. When connected to mnemory with INSTRUCTION_MODE=proactive (the default), the coding assistant will:

1. Load context at session start: Calls get_core_memories to load your technical preferences, project context, and recent activity.

2. Search before answering: When you ask about project architecture, past decisions, or "why did we do X?", the assistant searches memory for relevant context.

3. Remember important context: When you make architecture decisions, establish coding conventions, or share project context, it stores this automatically.

What Gets Remembered

The assistant will naturally store things like:

What	Type	Category	Example
Tech stack preferences	preference	technical	"Prefers TypeScript over JavaScript"
Coding conventions	preference	technical	"Uses single quotes, 2-space indent"
Architecture decisions	fact	project:myapp	"Chose PostgreSQL for the main DB"
Project structure	fact	project:myapp	"API lives in src/api/, uses Express"
Deployment setup	procedural	technical	"Deploy via GitHub Actions to AWS ECS"
Current sprint tasks	context	work	"Working on auth refactor this week"
Bug investigations	episodic	project:myapp	"Memory leak was caused by unclosed DB connections"

System Prompt (Optional)

Claude Code / Opencode don't require a custom system prompt — MCP server instructions handle everything. But if you want to add project-specific guidance:

When working in this project, use the project:myapp category for
project-specific memories. Key architecture decisions and coding
conventions should be stored as pinned facts.

Tips

• Use project:<name> categories: Tag project-specific memories with project:myapp, project:backend, etc. This makes searching by project easy.
• Pin architecture decisions: Important decisions (database choice, framework, deployment strategy) should be pinned so they load at every session start.
• Artifacts for deep analysis: When investigating a complex bug or doing architecture research, save the full analysis as an artifact. The memory holds the conclusion; the artifact holds the reasoning.
• Context type for current work: "Currently working on X" is a context type memory with 7-day TTL. It'll naturally expire when you move on.
• Shared vs agent-scoped: Technical preferences and project facts should be shared (no agent_id) so they're available across all your coding tools. Only agent-specific behavior rules need agent_id.

Multi-Project Setup

If you work on multiple projects, use category prefixes consistently:

project:frontend    — React app memories
project:backend     — API server memories
project:infra       — Infrastructure/DevOps memories
project:mobile      — Mobile app memories

The assistant can then search within a specific project:

search_memories(query="database schema", categories=["project:backend"])