Studied mem0's architecture by reimplementing it in Rust — sharing what I learned

[Adaimade]

Hey mem0 team and community,

I'm just getting started as an AI coding enthusiast — not a professional developer, but someone who loves exploring how things work under the hood.

I'm a big fan of mem0's memory architecture — the three-tier design (vector memory, graph memory, history) with the integer ID mapping trick and the ADD/UPDATE/DELETE/NONE dedup pipeline is really elegant.

As a learning exercise, I reimplemented the core memory logic in Rust (with the help of Claude Code), backed entirely by SQLite: R-Mem. It's intentionally minimal (~1,748 lines) — no plugin ecosystem, no multi-store support — just a study of how the core architecture works in a different language.

A few things I found interesting during the process:

• The integer ID mapping trick is brilliant for preventing LLM hallucination on UUIDs
• The contradiction-as-DELETE (not update) design choice makes the dedup pipeline much cleaner
• Graph entity extraction with self-reference resolution (I/me → user_id) is a subtle but important detail

Full credit to the mem0 team for the original design. Would love to hear feedback from the community — and happy to discuss the architecture with anyone interested!

[Adaimade]

Quick update — just shipped v0.2.0 with a few features that came directly from studying mem0's architecture:

• Tiered Archive: deleted/superseded memories are now preserved and searchable as fallback (something I wished mem0 had!)
• Anthropic native API: direct Claude support alongside OpenAI/Ollama
• FTS5 pre-filtering: full-text search before vector ranking for better recall
• MCP server: rustmem mcp exposes 7 tools for Claude Code / Cursor

Also got listed in awesome-sqlite. Still a learning project — would love any feedback on the approach!