Midas tests agent memory on BEAM with no LLM calls

Midas ran its first test on BEAM, a for in . It reached 0.56 recall@k on BEAM 100K and 0.51 on BEAM 500K. The test used 0 , spent $0 on API use, and sent no data out to an outside service.

The next planned tests are the 1 million and 10 million tiers. The goal is to keep improving Midas while staying and close to zero cost.

Key points

  • Midas reached 0.56 recall@k on BEAM 100K.
  • Midas reached 0.51 recall@k on BEAM 500K.
  • The test used no and no paid API use.
  • No data was sent out to an outside service.
  • Larger 1 million and 10 million tiers are planned next.
Read original