Solve Context Drift and Amnesia in Composer with Reponoesis (MCP Server)

[MathewAddala (ADDALA MATHEW)]

If you are using Cursor’s Composer or multi-file agentic edits, you have probably run into this issue:

Composer successfully modifies your backend model or config, but silently leaves your frontend client, documentation, or settings files inconsistent because it doesn’t trace the semantic connection between them. When you start a new composer run or open a new chat, the model has no context on why the previous changes were made.

I built Reponoesis to act as a local, cryptographic semantic memory graph for Cursor.

It runs locally on your machine via a standard stdio MCP server.

Key Benefits for Cursor Users:

1. Context Amnesia Prevention: When you start a new Composer chat, Cursor runs rpn_get_context to instantly load all active Architectural Decision Records (ADRs) and concept maps into memory.

2. Blast Radius Analysis: Cursor can run rpn_impact_map on files before writing edits, ensuring it understands the downstream consequences of mutating a file.

3. Deterministic Commit Gating: A local pre-commit hook (rpn check) blocks git commits if Cursor accidentally introduces logical contradictions or leaves files in a drifted state.

4. Agent Write-Back: Cursor uses the rpn_record_drift_explanation tool to write its own plain-English explanations of drift directly into the local SQLite database, replacing raw cryptographic warnings in the UI.

Quick Setup:

1. Clone the project:

   bash
   

git clone https://github.com/adithya/reponoesis.git

cd reponoesis

npm install && npm run build

cd packages/cli && npm link && cd ../mcp && npm link

2. Navigate to your project folder, run rpn init and rpn scan.

3. Add the server to your Cursor MCP settings:

   • Name: reponoesis

   • Type: command

   • Command: rpn-mcp --project /absolute/path/to/your/project

I’ve made the repository public on GitHub: MathewAddala/reponoesis

Would love to get your feedback on using this to make Cursor agent runs safer and more context-aware.