YourMemory: Biologically-Inspired Persistent AI Memory

YourMemory is a persistent memory layer for AI agents that uses biological forgetting curves to manage information decay and retention. It employs a hybrid vector-graph retrieval system to provide superior context recall compared to traditional vector-only databases. The tool is local-first, easy to install via pip, and compatible with major AI clients through the Model Context Protocol.

Key Points

• Implements the Ebbinghaus forgetting curve where memory strength decays over time but is reinforced by importance and frequent recall.
• Uses a hybrid retrieval system combining vector search and graph expansion to achieve 59% recall on the LoCoMo benchmark, significantly outperforming Zep Cloud.
• Provides a local-first architecture using DuckDB and local embeddings, ensuring privacy and zero-infrastructure setup.
• Supports multi-agent memory isolation, allowing different agents to maintain private contexts while sharing specific information.
• Integrates seamlessly with major AI clients through the Model Context Protocol (MCP) using simple tools like recall_memory and store_memory.

Sentiment

The community is predominantly skeptical of AI memory systems in general, not just this specific project. The most upvoted perspective argues that documentation files are superior to any memory layer. While some commenters defend the biological decay approach and distinguish between naive flat memory and more sophisticated typed systems, the overall tone suggests that the AI community has grown fatigued with memory solutions that add complexity without clear benefits. The project itself is treated respectfully as research, but the practical value proposition is widely questioned.

In Agreement

• Typed memory with different decay curves per category (personality vs. intent vs. emotion) solves the cross-project contamination problem that flat memory creates
• The Ebbinghaus forgetting curve is a legitimate model for managing context, and forgetting is an important feature for keeping agents focused
• Human forgetting is an evolutionary advantage, not a limitation — perfect recall would be suboptimal for prioritizing relevant recent information
• Spaced repetition and biological decay offer a principled approach to context management that outperforms static vector stores

Opposed

• Memory makes agents dumber — good markdown documentation (dev docs, roadmaps, checklists) solves every problem people think memory will solve, with no overhead
• The 'biological memory' framing is marketing fluff over basic cache mechanisms, and the benchmark claims are questionable
• All current memory implementations add overhead without demonstrably improving task performance, quality, or reliability
• Persistent memory causes agents to second-guess current tasks based on irrelevant prior conversations and comingle unrelated projects
• Everyone is building memory systems simultaneously yet no one has achieved a clean breakthrough, suggesting this is a fundamentally hard problem