• Category: Agentic coding
  • Status: discussion
  • Sources: 12gramsofcarbon post, HN discussion
  • Summary: A practitioner write-up (theahura, 2026-07-02) reports that giving coding agents searchable access to prior session transcripts produced no measurable performance gain over months of testing. The author argues agents cannot meaningfully curate or delete their own memory, treat all retrieved context as intentional, and accumulate scratch-pad noise that causes intent drift. The author's company accepts under 20 percent of automatically proposed memory updates and keeps a human diff review in the acceptance path, concluding that curated artifacts such as commit messages and documentation outperform raw transcript recall.
  • Why it matters: It pushes back on transcript indexing and automatic agent memory as a default, arguing that curated, human-reviewed memory beats replaying whole session histories.

Send feedback on this story