Roampal MCP
Roampal gives Claude Code and OpenCode persistent memory where advice that worked rises and advice that failed sinks.
- 54 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Hooks store your conversations with the agent on disk and feed them back into context.
- For OpenCode scoring runs as a separate call to a chosen model.
Install
- Works in
Manual install
pip install roampalThe first Quick Start command from the README. Then `roampal init` finds installed tools and configures them.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Hooks store your conversations with the agent on disk and feed them back into context.
- For OpenCode scoring runs as a separate call to a chosen model.
This is third-party code. Review the repository files before installing.
What it does
Roampal is a memory MCP server for an AI coding assistant that tracks the outcome of each answer. Before a reply it injects matching memories into context, after the reply it stores the exchange, then scores whether the advice helped. Good memories gain weight and bad ones lose it. For Claude Code this runs through hooks and a `score_memories` tool, for OpenCode through a plugin with an independent scoring model that stays off until configured. Memory is stored locally and computed on the CPU, with no GPU needed. The author reports 85.8% on a corrected LoCoMo benchmark.
Who it is for. People who have used Claude Code or OpenCode for months and are tired of re-explaining project conventions to the agent.
Good fit when
- An agent repeats the same mistake from session to session.
- One assistant serves several projects with separate memory profiles.
- You want successful approaches to surface on their own with no manual notes.
- Memory must stay local with no cloud database and no GPU.
Not a fit when
- Tasks are one-off and cross-session memory is not needed.
- Conversation records cannot be stored on disk.
- Ready memory for Cursor with no extra setup is expected.
Example request
Recall how we fixed the tests in this project last time and score whether it helped.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
Python 3.10 or newer and the `roampal init` command, which finds installed tools, are required. For OpenCode scoring is a separate call to a chosen model and stays off until configured. Exchange records and scores sit on disk, so account for that in private projects. README results come from the author's own single-GPU benchmark run with no independent check. Version 0.6.0 binds the memory profile to a directory.
How to disable. Remove the package with `pip uninstall roampal`, drop the hooks and the MCP entry from Claude Code settings or the plugin from OpenCode. Clear the memory data folder by hand.
FAQ
How do I connect Roampal?
Run `pip install roampal`, then `roampal init`. It finds Claude Code and OpenCode on its own. The `--claude-code` and `--opencode` flags pick one tool.
How does Roampal score memory?
In Claude Code the main model calls `score_memories` after an exchange. In OpenCode a separate observer model scores it and must be configured.
Is a GPU needed?
No. Roampal itself runs on the CPU. The author needed a GPU only to run benchmarks.
Does it work with Cursor?
The README lists Claude Code and OpenCode. Python hooks are described for Cursor too, but there is no ready `roampal init` setup for it.
Can I keep separate profiles?
Yes. A profile binds to a directory and the client passes it in the `X-Roampal-Profile` header.
Discussion
No questions or reviews yet.
Related
MCP serversThe Supermemory plugin for Claude Code: persistent memory across sessions with automatic prompt recall, conversation capture and a codebase index
Memory for AI agents with three types: facts, events and procedures that learn from failures; cloud by key or a local folder
MemoMind MCP gives Claude Code project memory, with stored facts, events and conclusions visible in a local dashboard.
Memory for Claude Code: hooks record human and agent decisions, and a grade is assigned afterwards once the outcome is known.
A Claude Code skill that pulls memory from Memori Cloud before every non-trivial answer and writes new facts back to it
Local memory for agents: stores conversations and project files verbatim and retrieves them with semantic search over MCP
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Roampal (this card) | no | not required | 54 | |
| agentmemory | yes | not required | 30,340 | stores agent memory; Roampal scores the outcome of each advice and lifts useful memories |
| MCP Memory Service | no | not required | 2,117 | an MCP memory service; Roampal adds Claude Code hooks and an OpenCode plugin with outcome scoring |
README in short
Outcome scoring turns plain memory into selection: advice followed by successful work gets priority. Hooks and the plugin do it with no user commands.