Memro, a new persistent memory system designed for AI coding agents, has emerged on Hacker News, offering a solution to the context amnesia plaguing tools like Claude Code, Cursor, and OpenAI Codex. Built in Go and backed by PostgreSQL with pgvector, Memro allows agents to share context across sessions via the Model Context Protocol (MCP). The project is currently source-available under the PolyForm Noncommercial License, with commercial use requiring a subscription to Memro Cloud or an enterprise license.
Architecture and Technical Implementation
The core of Memro is a Go HTTP API that leverages PostgreSQL 16 and pgvector for hybrid retrieval. It combines vector cosine similarity with full-text keyword search using Reciprocal Rank Fusion (RRF) with a k-value of 60. Agents authenticate using sovereign Ed25519 keypairs, where private keys are solved via local proof-of-work and never leave the host, or through standard Bearer API tokens. This architecture ensures that memory storage and identity management are decoupled from the agent itself, allowing for true persistence across restarts.
Integration and Developer Experience
Memro provides official SDKs for Python, JavaScript/TypeScript, and Go, alongside a CLI connector that can be initialized with npx memro-cli connect. The quickstart process is streamlined: developers connect their installed agent tools, verify status, and then use natural language commands like "Remember that our primary database is PostgreSQL 16 with pgvector." Upon restarting the session, the agent can recall this specific configuration detail. The repository includes runnable examples for Claude Code, Cursor, and Codex, demonstrating the integration in less than ten lines of code.
Advanced Memory Features and Verification
Beyond basic vector search, Memro introduces structured primitives such as Typed Memories (semantic, episodic, procedural, and profile), Temporal Search for tracking belief evolution, and Conflict Analysis to resolve contradictory facts across sessions. The project also emphasizes measurable verification, publishing evaluations on the LongMemEval benchmark across 500 questions. These tests focus on multi-session recall and knowledge updates, addressing common failure modes like stale facts and distractor rejection.
Key Takeaways
- Memro uses a Go backend with PostgreSQL and pgvector for hybrid search via Reciprocal Rank Fusion.
- Agents authenticate with sovereign Ed25519 keypairs, keeping private keys local.
- The tool supports Claude Code, Cursor, and Codex via MCP and official SDKs.
- Commercial use requires a Memro Cloud subscription or enterprise license; non-commercial use is free.
- Advanced features include temporal search, conflict resolution, and reasoning graphs.
The Bottom Line
Memro addresses a critical bottleneck in agentic workflows by treating memory as a shared, verifiable infrastructure rather than a proprietary feature of individual models. Its hybrid retrieval approach and focus on conflict resolution suggest a mature understanding of the limitations of simple vector stores in long-running coding sessions.