Retrieving the right memory can still make an agent worse.

Most memory-augmented agents paste a semantically similar past experience into the current context， but similarity does not mean the old advice fits the agent's present state.

Retrieval finds what looks similar； MemHarness asks whether any of it still applies now.

MemHarness adds a decision step between retrieval and action： the same policy compares the memory's original state with the current one， rewrites the transferable part， or rejects it and reasons without memory.

That reconstruction is learned end to end with Group Relative Policy Optimization from task rewards， rather than supervised labels for the rewritten guidance.

Using Qwen2.5-7B-Instruct， the method reaches 85.2% success on ALFWorld and 75.6% on WebShop， versus 76.4% and 66.1% for reinforcement learning without memory.

Removing reconstruction drops ALFWorld to 79.6%， while injecting raw memory reaches only 70.1%， showing that retrieved experience can become noise when the state has changed.

More unusually， disabling memory at test time still leaves the trained policy at 83.0%， above the 76.4% reinforcement-learning baseline.

Within these two interactive benchmarks， memory reconstruction appears to do two jobs： adapt past experience during inference and train the policy to judge situations more carefully even when no memory is available.

• arxiv. org/abs/2607.28272

Title： "MemHarness： Memory Is Reconstructed， Not Replayed"