{"id":46926968,"title":"Show HN: SuperLocalMemory – AI memory that stays on your machine, forever free","points":1,"user":"varunpratap369","time":1770493322,"time_ago":"11 hours ago","type":"link","content":"

*First open-source project. Built after getting tired of Mem0's $50/month bills.*

For the past year, I've been using Claude for everything—debugging, architecture reviews, code generation. The problem? Every new session starts from zero. "Remember that auth bug we fixed?" → "I don't have access to previous conversations."

I tried Mem0. Great product, but $50/month for cloud storage felt wrong when my machine has 1TB of disk sitting idle. I tried building my own with simple key-value storage. Worked okay for 100 memories, broke down at 500+.

So I spent 6 months implementing 2026 research: PageIndex (Meta), GraphRAG (Microsoft), xMemory (Stanford). The result is SuperLocalMemory V2.

*Architecture (what makes it different):*

- *Layer 1:* SQLite + FTS5 for raw storage (<10ms writes)\n- *Layer 2:* Tree hierarchy for context (parent-child relationships)\n- *Layer 3:* Knowledge graph with auto-clustering (TF-IDF + Leiden algorithm)\n- *Layer 4:* Pattern learning (learns your coding style: "You prefer React over Vue - 73% confidence")

*It works everywhere:*\n- MCP integration → Auto-configures Claude Desktop, Cursor, Windsurf, Continue.dev\n- Universal skills → Works in Claude Code, Cody, VS Code extensions\n- CLI → `slm remember "content"` from any terminal

*What I learned building this:*

1. *Graph clustering is expensive* - Leiden algorithm is O(n²) worst case. Solution: profile splitting for >1000 memories.

2. *Backward compatibility is hard* - V1 users upgrading to V2 hit schema changes. Solution: Try-catch on every ALTER TABLE with existence checks.

3. *MCP auto-configuration is fragile* - 11 different IDEs, 11 different JSON formats. Solution: Template system with jq fallback.

*Current status:* 2.2.0 release. 100% test coverage (35/35 passing). Works on Mac/Linux/Windows. Single `./install.sh` script.

*Performance (tested up to 10K memories):*\n- Add memory: <10ms\n- Hybrid search: 80ms (combines semantic + exact + graph)\n- Graph build (100 memories): <2s\n- Storage: 60-96% compression via tiering

*When NOT to use this:*

- If you need cloud sync across devices → Use Mem0 or Zep\n- If you need enterprise SSO/audit logs → Use Zep Enterprise\n- If you don't code → This is developer-focused

*GitHub:* https://github.com/varun369/SuperLocalMemoryV2\n*Install:* `git clone https://github.com/varun369/SuperLocalMemoryV2.git && cd SuperLocalMemoryV2 && ./install.sh`

*Honest ask:* This is my first open-source project. I'd love feedback on:\n- Architecture decisions (especially the 4-layer approach)\n- Performance bottlenecks (graph build feels slow)\n- Integration priorities (what tools to support next)\n- Documentation clarity

I'm a senior manager at Accenture (25+ years) with basic coding skills. Built this because I needed it, not because I wanted to compete with Mem0/Zep (they're great products).

*What's next:* V3 release with `npm install -g superlocalmemory` for easier installation. Same features, better distribution.","url":"https://github.com/varun369/SuperLocalMemoryV2","domain":"github.com","comments":[],"comments_count":0}