PRO-LONG beats agent summarization by 18 points using grep

PRO-LONG skips the summarization step every agent-memory system leans on, storing the full log instead and letting the agent grep it, per the paper.

That matters for anyone building agent memory. PRO-LONG beats a base coding agent by 18.0 percentage points across frontier models on the full ARC-AGI-3 public game set. It hits up to 76.1% pass@1 while spending 4.2 to 5.8 times fewer tokens than a summarize-and-evict setup. Run with Fable 5, it reaches 97.4% best@2 for $1,750 total. The code is public on GitHub.

The logic holds up once you sit with it. Summarization is lossy compression done before you know what you'll need. You're guessing at write time what'll matter at query time. Grep is lossless storage with retrieval deferred until the agent actually needs the answer. That's also where the token savings come from. You're not paying to re-read a summary every turn. You're paying only for the three lines you pulled out.

Grep has a number, the other idea has a manifesto

This landed the same week a senior Anthropic engineer's 12-page treatment of graph engineering went viral, four threads in 48 hours. It argues agent state needs an explicit typed node graph instead of a log. Andrew Ng's team shipped a course on the same idea days later. Turing Post is already running a piece asking if the whole thing is real. A definitive-guide post and a skeptic piece landing in the same stretch is what a hype phase looks like, not a settled one. PRO-LONG skipped that step and shipped a benchmark instead. A related paper, ContextSniper, points the same direction. Filtering verbose tool output cut Claude Code's tokens 38.9% on SWE-bench Lite tasks with resolution rates essentially unchanged.