GitHub project YourMemory introduces a biologically inspired memory layer for AI agents, replacing static storage with an Ebbinghaus forgetting curve. The tool, written in Python and backed by DuckDB, assigns each memory a strength score that decays unless reinforced by recall. By pruning low‑strength items every 24 hours, the system keeps context relevant and token‑efficient for developers and in the future tool.

Benchmarking against the LoCoMo dataset, YourMemory achieved 52 % Recall@5—almost double the 28 % of the Zep Cloud baseline—while cutting token waste by roughly 84 %. The hybrid retrieval engine first pulls top matches via cosine similarity, then expands through a graph traversal to surface semantically related memories that traditional vector search misses and improve reasoning in long running conversations by maintaining context.

YourMemory ships as a lightweight MCP server, requiring only a single pip install and no external database. Integration with Claude, Cline, or any MCP‑compatible client is automated through configuration files, enabling developers to add a recall‑store‑update workflow with a single sample instruction set. The project demonstrates that biological decay can make agent memory scalable and cost‑effective for enterprise applications today.