past.dev launches a long-term memory API for AI agents

Summary

past.dev presents a long-term memory API designed to help AI agents maintain accurate, changing facts across emails, meetings, chats, documents and other large histories. Developers ingest text with its original timestamp, then call a recall endpoint with a question and requester identity; the service returns ranked evidence, dated sources, an as-of time and token usage. It tracks current and previous values, so a later record can supersede an earlier one, and supports audience filtering so reads expose only memory available to the requesting identity. The API also includes ingestion-status and deletion endpoints, while MCP clients such as Claude, Cursor and VS Code can connect through an authenticated endpoint. The company says its system reached 92.08% at 100K tokens, 90.65% at 1M and 85.03% at 10M on the BEAM benchmark, ranking first among published results at each listed history size. Its published harness also evaluates BEAM and corrected-key LoCoMo results, although the page does not provide the full comparison table. In a separate, non-official DolphinBench measurement of 599 paired tasks from three users, Claude Code with past.dev completed 519 tasks correctly at an average cost of $0.13, compared with 323 correct tasks at $0.27 without it; the page attributes grading to DolphinBench's Claude-based judge and says the runs used different memory configurations. past.dev offers a managed service or enterprise deployment in the customer's AWS, Google Cloud or Azure account, including database-only or full deployment options. The page lists SOC 2 Type II, ISO 27001 and ISO 27701 claims, and advertises pay-as-you-go access starting with 150,000 credits for $45, a $99 monthly Flex plan with 400,000 credits and unlimited recall, and custom enterprise plans from 5 million credits per month. New users sign up and create a project key in the console.