Contrary to intuition, agents gain no performance advantage from accessing past session transcripts for software engineering tasks, even with extensive search capabilities. Researchers found that providing agents with other forms of context sufficiently covers their needs. The notion of session transcripts as a rich source of information, akin to 'new oil,' appears to be a misconception, as built-in context already distills essential details.

This finding challenges the common practice of storing and indexing entire transcripts for agent recall. Architects often deploy vector or SQL search layers over these logs, assuming it augments agent understanding. However, tests showed this additional processing often leads to models performing worse. Agents end up re-processing known information and consuming valuable tokens on irrelevant scratchpad data.

Models struggle with context pruning, a vital trait for long-term memory. Without state, agents treat all input as ground truth, leading to compounding intent drift. This issue is exacerbated because coding benchmarks penalize models for assuming incorrect input. Consequently, automatic transcript indexing becomes a liability, bloating bills and degrading model quality rather than improving agent effectiveness.