JitMem skips memory summaries, beats baselines by 16.2 points

Most agent memory systems compress experience as it happens. An agent finishes a task, writes a reflection or extracts a reusable skill, then throws away the raw log. JitMem does the opposite. It keeps the raw trajectories and waits.

That matters for anyone building agents that run long enough to accumulate real history. Write-time compression forces a bet on what future tasks will need, made before those tasks exist.

When a new task arrives, a curator assembles a task-specific memory payload from the stored logs, built for that task alone. The curator can train directly against whether the assembled memory helped the agent succeed, since it knows the target before it curates.

The numbers back the bet. JitMem beat the strongest baseline by 16.2 points on ALFWorld, 16.3 on WebShop, and 3.9 on tau-squared-bench, according to the paper posted to arXiv. Those are three different task families: agents navigating simulated homes, agents shopping, agents handling multi-turn tool calls. The gap holds across all three.

What the paper doesn't settle is storage cost. Keeping every raw trajectory instead of a distilled summary means memory grows with every task the agent runs. Nothing in the writeup says how that scales past benchmark-sized task sets, or what retrieval looks like once the log spans a year of production traffic instead of a few hundred episodes.

The fix for that gap probably isn't smaller logs. It's a curator that can search a large raw store fast enough to stay task-specific instead of falling back to write-time summaries out of necessity.