MCP Memory ships on GitHub - persistent agent memory becomes a commodity primitive

A new open-source MCP server implements persistent agent memory with SQLite FTS5 and Google's OKF—no external API, no vector database to manage, and compatible with any MCP client. The first infrastructure gap in agentic pipelines has just been commoditized.

In plain terms: Agents forget everything between sessions unless you build the memory layer. MCP Memory wraps SQLite with a Model Context Protocol server - functional, local, zero external dependencies.

The fact

MCP Memory (github.com/fellowgeek/mcp-memory, surfaced on Hacker News) implements persistent agent memory as a MCP server using SQLite FTS5 for retrieval and Google's OKF for embedding storage. Compatible with any MCP client - Claude Code, Cursor, and the growing ecosystem. No external API required; memory lives in a local SQLite file.

Our read

The first infrastructure gap developers hit with agentic pipelines is statelessness—the agent doesn’t remember session N when starting session N+1. Solutions so far have required either a vector database (operational overhead, external dependency) or custom database plumbing. MCP Memory takes the trade the right way for most use cases: FTS5 BM25 ranking instead of dense vector retrieval gives slightly lower recall quality but eliminates the infrastructure layer entirely. For privacy-sensitive enterprise deployments, local storage is a feature, not a compromise. The pattern (SQLite-backed MCP memory server) is simple enough that it will be absorbed into major agentic frameworks as a default in the next framework release cycle—6–12 months.

[Under the hood] FTS5 provides BM25-ranked full-text search over memory chunks. Google's OKF handles embedding storage for semantic similarity. Local file = portable, backupable, auditable—no cloud dependency.

Watch for

Whether Claude Code, Cursor, or Windsurf adopt a similar pattern natively—the moment one major client ships built-in persistent memory, standalone MCP memory servers become optional rather than necessary.