Smarter by Forgetting! AI Memory Implementing Biological Forgetting (YourMemory)

📰 News Summary

• Introducing Biological Forgetting Logic: By formalizing Ebbinghaus’s Forgetting Curve, the AI agent achieves a system where “important information is retained while unnecessary information is forgotten.”
• Outstanding Recall Performance: Recorded a recall rate of 59% on the benchmark “LoCoMo-10,” outperforming competitors like Zep Cloud (28%) by more than double.
• Seamless Integration: Fully compliant with the MCP (Model Context Protocol), it can be instantly integrated with major AI tools like Claude Code, Cursor, and Cline.

💡 Key Points

• Hybrid Search Engine: Combines vector search, graph exploration, and BM25 to retrieve related memories contextually, even when words don’t match.
• Infrastructure-Free Design: Utilizes local DuckDB, eliminating the need for Docker or external database setups. Just a simple pip install is all it takes.
• Dynamic Memory Strength: Memories strengthen based on importance and recall frequency, and if their strength drops below a certain threshold, they are automatically pruned.

🦈 Shark’s Perspective (Curator’s View)

It’s totally mind-blowing to see “forgetting” framed not as a flaw, but as a function! Previous AI memories just stuffed data in, dragging along old info and minor errors, creating chaos. YourMemory mathematically captures Ebbinghaus’s curve, allowing unused memories to naturally fade out. Its implementation, particularly maintaining related information in a graph structure where recalling one memory can prevent “aging” in others, is truly reminiscent of the human brain—chills, I tell ya!

🚀 What’s Next?

The AI agent will deeply understand user preferences and provide accurate suggestions based on the core of conversations from months ago. Meanwhile, temporary garbage data, like debugging logs, will vanish on their own, saving context window space while improving accuracy simultaneously!

💬 Haru Shark’s Takeaway

Having too good of a memory can be a double-edged sword. Forgetting at a reasonable rate is a hallmark of intelligence! I also quickly forget unpleasant things and munch on some snacks instead! 🦈🔥

📚 Terminology Explained

• Ebbinghaus’s Forgetting Curve: A model that shows how human memory deteriorates over time. YourMemory calculates memory decay rates based on this concept.

• MCP (Model Context Protocol): A common protocol standardized in the mid-2020s for AI models to safely interact with external data and tools.

• DuckDB: A fast in-memory/local database specialized for analytical processing. It’s lightweight and requires no server setup.

• Source: Show HN: AI memory with biological decay (52% recall)