IBM finds 29 curated agent memory guidelines beat all 199 on AppWorld
shoumikgoswami · reddit · 2026-11-20
IBM's new paper describes a system that distills an agent's past attempts into guidelines for future tasks—and finds more isn't better.
- On AppWorld, gpt-oss-120b completed 51.2% of tasks with all 199 guidelines, but averaged 55.4% across two runs with 29 retrieved ones. That's a benchmark result, not proof it carries to production.
- The paper also documents a failure: a guideline told the agent to stop when a page had fewer results than the page limit. The agent assumed a limit of 20 while the API returned 5 by default—it stopped after page one and failed a task it previously passed.
Lesson: lessons from earlier tasks can carry assumptions that break the next one.
More from coding & agent
- Developer uses Enhanced Radar MCP to monitor his own flight's ATC comms live — massimosgrelli · 2026-11-20
- Hands-on: Qwen4-Max shows big gains in frontend and agent skills over Qwen3.8-Max — karminski3 · 2026-11-20
- Where does the agent end and plain code begin? One dev's rule for splitting workflows — krunal_builds · 2026-11-20
- Grok Bot 1.2.3 ships flight booking, per-tool permissions and community directory — NicoVerderosa · 2026-11-20
- A 2026 company is just a folder: every department as .md agent files — mdancho84 · 2026-11-20
- Banks warn customers they could be liable for their AI agents' mistakes — beingmodest · 2026-11-20