AI memory system uses biological decay to improve recall
The innovative AI memory system, YourMemory, leverages the Ebbinghaus forgetting curve to function as a dynamic context substrate, attaining a Recall@5 rate of 52% on the LoCoMo dataset. Each memory is given a strength score, where recall enhances data and reduces decay curves, while unused information is eliminated. To address the logical neighbor issue, a graph layer is implemented over the vector store, resulting in an approximate 84% reduction in token waste. Designed as a local-first MCP server utilizing DuckDB, this system is aimed at supporting long-duration agent projects.
Key facts
- Uses Ebbinghaus forgetting curve for memory decay
- Achieves 52% Recall@5 on LoCoMo dataset
- Cuts token waste by roughly 84%
- Graph layer over vector store solves logical neighbor problem
- Built as local-first MCP server using DuckDB
- Targets long-running agent projects
- Memory strength score with spaced repetition
- Unused data pruned when threshold reached
Entities
—