Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-25 · high
Facts kept only in conversation vanished at the first compaction and stayed missing through 106 of 108 later summaries.
Why now: As of July 25, the findings run counter to the assumption behind most agent memory tools: that they'll get used voluntarily.
A coding agent made zero memory calls across 114 turns despite a pre-seeded memory store, per a new evaluation from researcher Swapnanil Saha. That's a problem for any harness built around a recall tool. The design assumes the agent will ask for a fact when it needs one, and this run never did.
Saha's argument: memory can't be a document an agent opens on its own. It has to be a harness property, delivered whether the agent asks or not.
Saha ran a decay probe with ten facts held only in conversation, no harness backup. Every one vanished at the first compaction summary and stayed missing through 106 of 108 later compact-resumes.
The deprived agent's workaround was grepping the harness's own session files to rebuild what it had lost.
The fix that worked was deterministic, cue-triggered injection: facts pushed into context on a trigger instead of fetched by choice. That delivered in every seeded run with zero false alarms, and harness-injected facts survived all 138 compact-resumes tested.
Injection isn't free. Saha also measured that 39% of intra-session re-reads paid again for content the agent already had before a compaction boundary. That's redundant spend even when the memory system works exactly as designed.
Anyone wiring memory into a harness should design for push, not pull.
Each link below shares sources, entities, or timing with this story.
A June 16 position paper argues today's benchmarks predate AI agents: they conflate multiple system components into single scores, penalize valid alternative solutions, and lack the granular feedback needed to iterate on agent systems. Read the current wave of open-weight SWE-...
Traced across 557 SWE-chat sessions (94,813 events) and 33,097 agentic pull requests from AIDev. Agent-facing artifacts account for 60.5% of documentation interactions versus 10.6% for classical technical docs and 1.3% for API references. Consultation is self-initiated 70.2% o...
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...
Logistic-regression probes on a coding agent's hidden states can decode whether code will parse and pass tests at AUC up to 0.83, and those internal representations run ahead of the agent's own edits, predicting outcomes as much as 25 steps in advance (arXiv). The authors call...
HarnessOpt-Bench (arXiv 2608.06301) has a frontier LLM act as an optimizer receiving a target agent's seed harness (prompts, tools, control flow, memory, orchestration code) plus graded eval feedback and a fixed evaluation budget, then edits and nominates a candidate scored on...
CodeGrep measures a 30B OpenHands agent averaging 23 rounds and 631K tokens per resolved SWE-Bench Verified issue, much of it grep, glob and view_file. A 14B retrieval agent trained end-to-end with GRPO raises resolve rate to 27.0% from 25.8% while cutting 15% of rounds and 19...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.