Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-21 · high
Terminal-Bench 2.0 scores jumped 8.3 points when timing beat retrieval on injecting reminders.
Why now: Covered in the July 21 briefing on long-run agent decay.
A second agent decides when to inject a reminder into a first agent's long task, lifting Terminal-Bench 2.0 scores 8.3 points, per the arXiv paper 'Remember When It Matters' (2607.08716).
That's a rare hard number on a real failure mode: agents burying or dropping instructions given at the start once a task runs long. Anyone running unattended agents past a few dozen steps, coding agents, support loops, or ops automation, has probably hit this already.
One agent does the task. A second, dedicated memory agent watches it work.
No fixed timer. No passive bank waiting to be queried. Reminders land only at the moment they matter.
The gain held on a second benchmark, τ²-Bench, at 6.8 percentage points. It showed up for both weak and strong action agents, evidence the fix targets a structural gap, not a weak model's shortfall.
The paper's own comparison is the interesting part: selective, timed injection beat both a passive memory bank and always-on injection. I'd have bet on always-on being safer. It lost anyway, apparently because constant reminders bury the one that matters under noise.
It bolts onto an existing agent framework rather than replacing it, which is good news if you've already got one built. I want to know what running a second agent watching the first costs in tokens or latency, something the paper doesn't say.
Each link below shares sources, entities, or timing with this story.
Instead of cloning full teacher demonstrations that mismatch the contexts a student hits at test time, spend a fixed teacher-labeling budget on short continuation rollouts that branch from the student's own trajectories (arXiv). On HotpotQA, ALFWorld, and Terminal-Bench-Dev, b...
Here's the number that should sit in every "agents will replace engineers" thread: 15.2%. That's the best model. The mean across 15 frontier models is 4.3%. Tencent Hunyuan's Long-Horizon-Terminal-Bench put 15 frontier models against 46 long-horizon terminal tasks across nine...
arXiv 2607.26598 targets the failure where an agent recovers from an error within an episode but hits the identical failure in later tasks, because post-episode feedback never revises the persistent harness. Guided by a domain-level Evolution-SOP, it writes episodic memory rec...
Same weights. Same tasks. Same time budget. One change to the harness, and fail-to-pass on SWE-bench Verified went from 28% to 49%. The setup in arXiv 2608.26218 is clean enough to be worth describing precisely. 169 SWE-bench Verified tasks, a 20,480-token context window, a fi...
arXiv 2608.08311 describes an agent that continuously rewrites its own tools, prompts, context assembly, and core implementation through reviewed commits. On Opus 5: 86.74% Terminal-Bench 2.1, 90.69% OSWorld-Verified, 0.2301 normalized reward on CL-Bench. It also documents "Ho...
Shang et al. show self-evolving agents improve only up to a critical skill-pool size, after which new skills actively hurt. The mechanism: a defective skill becomes reference material for distilling later skills, forming cross-round contamination chains. Deleting the source sk...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.