DEV Community

News
News

Posted on Edited on

Memro: a memory layer for AI agents with typed memory and MCP over PostgreSQL

Infographic cover

๐Ÿ“… 14 November 2026 โ€ข memrohq/memro Memro is a memory layer for AI agents. One public repository at github.com/memrohq/memro ships a Go HTTP API, official SDKs for Python, JavaScript and Go, an MCP server over stdio and SSE HTTP, and a CLI. The release landed on November 14, 2026, shown as Show HN on the same day, and the repository itself was created on October 18, 2026. It is one of several approaches to the agent-memory problem, not a settled answer.

Storage, search and licence. The backend runs on PostgreSQL 16 with the pgvector extension, so vectors and full-text records live in the same database. Hybrid search merges vector cosine similarity with full-text search through Reciprocal Rank Fusion (RRF) at k=60. The documentation for that hybrid path expects an OpenAI API key, set separately in an environment file. The code is source-available under the PolyForm Noncommercial License 1.0.0: free for non-commercial use, while commercial use requires Memro Cloud or a corporate licence.

Four kinds of memory. Memro separates semantic memory for facts, episodic memory for events and conversations, procedural memory for instructions and recipes, and profile memory for agent constraints. On top of that sit temporal search, tracking of how beliefs evolve, conflict analysis through a resolve_conflict function, and reasoning graphs that trace where an observation came from, which causes are linked, and what depends on what. The material covering the release names MemGit and Hippocampus, other memory systems for Claude Code, as competing approaches.

Identity and packaging. An agent authenticates with a sovereign Ed25519 key pair and a local proof-of-work step, so the private key never leaves the host. Standard Bearer API tokens remain an alternative, and the official SDKs sign requests automatically. Release v0.1.0 publishes pinned images ghcr.io/memrohq/memro:0.1.0 and ghcr.io/memrohq/memro-dashboard:0.1.0, verifiable with Cosign. The local stand brings up the API on port 8081, the dashboard on 5173 and PostgreSQL on 55432.

Tools and maturity. The MCP surface exposes five tools: remember, recall, get_recent_memories, delete_memory and export_memories, so any MCP-capable client can read and write memory. A single npx memro-cli connect command auto-configures Claude Code, Cursor, Codex, VS Code, Claude Desktop, Antigravity and Windsurf to the hosted MCP server. At the time of checking the repository had 6 stars, 1 fork and 0 open issues, so the project is at an early stage rather than a proven one.

Sources: github.com, memro.co, wpnews.pro, emergent.sh

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)