roampal-core

Overview

Roampal-core is an innovative MCP server that implements outcome-based persistent memory for AI coding assistants like Claude Code and OpenCode. It automatically injects relevant context before AI responses and scores outcomes to promote good advice while demoting bad ones. The system maintains five memory collections with different lifespans: working (24h), history (30 days), patterns (persistent while useful), memory_bank (permanent), and books (permanent reference docs). This creates a feedback loop that makes the AI assistant progressively smarter with each exchange.

Try asking AI

After installing, here are 6 things you can ask your AI assistant:

• search_memory

  Deep search across all memory collections

  Try:Find previous solutions to authentication issuesTry:Search for information about JWT refresh patterns
• add_to_memory_bank

  Store permanent facts (identity, preferences, goals)

  Try:Remember that I prefer never staging git changesTry:Note that my project deadline is June 30th
• update_memory

  Correct or update existing memories

  Try:Update my preference to use TypeScript instead of JavaScriptTry:Correct my address in the memory bank
• delete_memory

  Remove outdated information from memory

  Try:Delete the old auth solution from last monthTry:Remove my discontinued preference for manual testing
• score_memories

  Score previous exchange outcomes (Claude Code only)

  Try:Rate how helpful the last response wasTry:Score the quality of the solution provided
• record_response

  Store key takeaways from significant exchanges

  Try:Remember that the auth bug was caused by expired tokensTry:Note that the database solution worked well

Comparable tools

Installation

pip install roampal
roampal init

roampal init --claude-code
roampal init --opencode

{
  "mcpServers": {
    "roampal": {
      "command": "roampal",
      "args": ["start"]
    }
  }
}

roampal start  # Start the HTTP server manually
roampal status  # Check if server is running

FAQ

How does Roampal improve AI performance over time?

Roampal scores each interaction outcome, promoting good advice and demoting bad advice. This creates a feedback loop where the system gets progressively smarter with each exchange, learning from past successes and failures.

What memory collections does Roampal maintain?

Roampal maintains five memory collections: working (24h current session), history (30 days of past conversations), patterns (proven solutions), memory_bank (identity, preferences, goals), and books (uploaded reference docs).

How does Roampal handle different platforms like Claude Code and OpenCode?

The core loop is identical across platforms, but the implementation differs: Claude Code uses hooks for context injection and capture, while OpenCode uses a plugin. OpenCode also uses an independent sidecar for scoring to avoid self-assessment bias.

On Hacker News

Recent discussion from the developer community.

• Show HN: I made Claude Code learn from its mistakes▲ 4
  Story by roampal · 2025-12-16

Compare roampal-core with