Structured Agent Memory Beats Trajectory Dumps by 10.7 Points

Query-Conditioned Reuse cut token use 48.9% and raised task success 10.7 points over trajectory-dump memory, per a paper posted to arXiv.

That's a design question for anyone building agent frameworks, not a tuning knob. Most systems still dump a full retrieved trajectory into context on every call, and QCR shows that pattern costs money and accuracy at the same time.

The dump approach degrades as traces get longer and as the values a stored trajectory relied on drift from the new task's specifics.

QCR replaces the dump with a structured memory: a reusable procedure plus the bindings, applicability conditions, and verification steps needed to reuse it safely.

Tested on WebArena, WorkArena, and AppWorld, about 2,400 target tasks, QCR averaged 62.3% success. Its reranker picked an appropriate stored memory 94.8% of the time, per the paper.

Both numbers move the same direction, which is rare. Cut tokens in most systems and accuracy drops. Prune too hard and the agent forgets what it needed. QCR moved efficiency and success the same way at once.

The reranker's hit rate matters more than the headline success number, since QCR only works if it picks the right procedure for a new query. The paper doesn't say what happens to accuracy when a picked memory's bindings don't transfer to the task at hand.