YourMemory

The software provides persistent, session‑spanning memory for AI agents that follow the MCP protocol. It stores information in a hybrid structure combining vector embeddings, a graph of relationships, and a biologically inspired decay model based on the Ebbinghaus forgetting curve. Retrieval, pruning, and graph expansion are performed locally without external LLM calls, allowing the system to keep a concise, up‑to‑date context while discarding stale facts.

Developers can register agents, designate memory as shared or private, and control visibility through a matrix of rules. The memory engine supports chain‑aware pruning and recall propagation, enabling multi‑agent scenarios where knowledge is shared or isolated as needed. Two simple commands let agents store and retrieve scoped memories.

Benchmarks on the public LoCoMo‑10 dataset show a 59 % Recall@5, outperforming comparable systems while reducing token usage and LLM calls across multiple sessions. The implementation is open source, self‑hostable, and runs entirely on the user’s hardware.