Most existing AI agents struggle with data decay, often relying on simple vector search or stuffing raw history into context windows—methods that frequently fail to distinguish between current facts and obsolete information. In the BEAM benchmark, which tests performance across massive datasets, models using traditional full-history retrieval topped out at 25.9% accuracy for million-token inputs.

past.dev sidesteps this by indexing data with specific timestamps and replacement logic. The API allows developers to query information while respecting access controls, ensuring that agents only retrieve facts authorized for a specific user. According to CEO Mehdi Djabri, the industry's reliance on similarity search has been a fundamental error, as search identifies what looks relevant rather than what remains true.

Performance on the BEAM benchmark, published at ICLR 2026, highlights a significant lead over existing solutions. At the 10-million-token mark, past.dev scored 85.03%, compared to the previous best of 68.0%. The company has released its evaluation harness to the public, allowing for independent verification of these results. The platform is currently available for self-serve sign-up, supports the Model Context Protocol (MCP), and adheres to SOC 2 Type II and ISO compliance standards.