Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

My AI coding agents read 1.2 billion tokens in 32 sessions. Most of it was text they'd already seen

Description and use

Simem cuts token usage by providing a concise project primer, sharing memory across apps, and filtering noisy tool output, saving up to 22 % of tokens.

Problem

AI coding agents repeatedly re-read the same project context and tool output, wasting tokens and time.

Outcome

Simem reduced token consumption from 1.16 M to 0.90 M across all tool reports (22% reduction) and from 26 k to 14 k for test reports (46% reduction), while ensuring no failure lines were dropped and memory persisted across sessions and apps.

Workflow

  • Install simem via two commands in Claude Code or Codex
  • Configure simem to use the project’s first commit as ID
  • On session start, simem supplies a short summary of recent work, decisions, and test commands
  • During interaction, simem appends relevant saved notes to each user message
  • When a user asks about past decisions, simem searches the saved memory
  • Simem trims tool reports to keep only failures and essential lines
  • Large file reads are guarded by outline first and re‑read flags

Evidence

about 1.2 billion tokens, across 32 work sessions in Claude Code and Codex.

Source

most of it wasn't new. It was the same conversation being re-read turn after turn.

Source

simem gives the agent a short summary of your project before you type anything.

Source

simem trims noisy tool output before it lands in the conversation: 81–99% off a full test or install report, and it never drops a failure line.

Source

All tool reports 1.16M tokens → 0.90M tokens, 22% shorter.

Source

  • Record date: 2026-10-23T03:11:40.534Z