Renor, a new open-source memory engine for AI coding agents, hit Hacker News today with a promise that cuts to the core of developer frustration: shared, persistent context across different AI tools. Built on the Model Context Protocol (MCP), Renor allows a single memory store to be accessed by Claude Code, Codex, Cursor, Grok, Gemini, and others. The project, created by a developer tired of re-explaining project decisions to forgetful agents, uses a plain Markdown ledger that users own and can export.

The Architecture of Forgetting No More

The core innovation here is simplicity. Instead of complex vector databases or proprietary cloud locks, Renor saves agent learnings as plain Markdown files in a local folder (defaulting to ~/.renor). SQLite handles the search index, which can be rebuilt from the Markdown at any time. This means no embedding model is required for basic keyword search, though hybrid search (keywords plus meaning) is available for self-hosters who opt in to OpenAI or local models. The package is lightweightβ€”1.8 MB to download, ~5 MB installed, and about 143 MB of RAM for 1,000 saved memories.

Benchmark Results and Token Efficiency

The developer claims good results with fewer tokens, backed by specific benchmarks from October 2026. On LongMemEval-S, Renor achieved 94.4% accuracy with default settings and 98.4% with 'deep search' enabled (top-K 50, 4,000-token budget). LoCoMo tests showed 96.7% accuracy. The tests used Claude Opus 5.5 for LongMemEval and LoCoMo, and DeepSeek V4.1 Flash for ConvoMem. These numbers suggest that a well-structured, simple retrieval system can outperform more complex setups in specific long-context memory tasks.

Integration and Privacy

Integration is frictionless for most users. A single command, npx renor connect --client [agent-name], adds Renor to the agent, handles authentication via the app's own flow, and verifies connectivity. For the privacy-conscious, self-hosting requires Node.js 22.16+ and keeps all memory local. Telemetry is on by default but anonymized (no text, names, or paths) and can be disabled with DO_NOT_TRACK=1. The engine is licensed under Apache-2.0, with the CLI and SDKs under MIT, signaling a commitment to open ecosystem growth.

Key Takeaways

  • Renor uses plain Markdown for memory, ensuring user ownership and easy export.
  • It supports major AI coding agents via the Model Context Protocol (MCP).
  • Benchmarks show high accuracy (94.4% on LongMemEval-S) with low resource overhead.
  • Self-hosting is possible with Node.js 22.16+, keeping all data local.
  • The project is Apache-2.0 licensed and encourages contributions via DCO.

The Bottom Line

Renor tackles the fragmentation of AI agent memory with a brutally simple, open-source solution. If it delivers on its benchmark promises without locking users into proprietary clouds, it could become the standard for persistent agent context.