A GitHub project has released an AI memory system inspired by biological decay mechanics to provide persistent recall for agents across conversations.

According to the primary source, the system integrates BM25, vectors, graphs and decay to reach 59% recall on a benchmark of 1,534 QA pairs from 10 multi-session conversations, doubling Zep Cloud's 28% (https://github.com/sachitrafa/YourMemory). It uses effective_λ = base_λ × (1 - importance × recall_factor) with category-based decay rates from 38 days for strategies to 11 days for failures.

This approach addresses shortcomings in existing memory tools that the original coverage did not fully explore, such as the tendency for vector stores to retain obsolete data leading to context pollution, a problem discussed in analyses of catastrophic forgetting (Kirkpatrick et al., PNAS, 2017, https://www.pnas.org/doi/10.1073/pnas.1611835114). Unlike cloud services, this requires no infrastructure and runs via local MCP server.

Synthesizing with Zep's methodology and Ebbinghaus's original experiments from 1885, the project charts a new direction for human-like AI by allowing memories to fade and be replaced organically, which could improve long-term agent efficiency and adaptability beyond what static RAG systems offer (https://www.getzep.com/).