MemCon lifts agent success 15 points, trims token cost

MemCon frames agent memory management as a Markov Decision Process, per the paper. An online policy decides when to retrieve a memory, reuse a saved plan, re-query after getting stuck, or consolidate and forget.

Across 6 benchmarks, 3 agent frameworks, and 3 model backbones, the policy lifted task success by up to 15.2 points, according to the paper. It also cut token use 5 to 20%. That's a rare pairing. Most memory fixes buy accuracy with more tokens, or save tokens by losing accuracy.

The policy isn't a trained model. It's a tabular contextual bandit with UCB exploration, and it only needs binary success or fail signals from finished tasks to update. No pretraining step, no labeled memory dataset.

The paper doesn't say how the policy holds up once an agent's memory store grows well past what these six benchmarks tested. It also doesn't say whether UCB exploration adds latency to retrieval decisions inside a live agent loop.

The no-pretraining part is the detail I'd weigh more than the 15-point headline number. A big accuracy jump is easy to lose once you change benchmarks or swap backbones. A bandit that learns from pass/fail signals and skips training is what actually survives being dropped into an already-running agent stack. It doesn't ask for a training set first.