A new system scrubs robot memories before an AI agent reuses them, instead of just fetching whatever looks similar.
Researchers describe Memory Adaptation for Task-Conditioned Execution, or MATE, a post-retrieval step that sits between an embodied agent's memory bank and its next move. Past approaches assumed that a trajectory which worked once, and matches the current task semantically, is good enough to replay. MATE instead strips out obsolete control context, pulls condition-action-effect transitions from the stored trajectory, normalizes the actions against what is actually verified to work, and repackages the result into a compact, task-specific format under a fixed token budget, with no extra LLM calls required. On 134 ALFWorld household-robot tasks, agents using MATE hit task success rates of 81.3 percent with Qwen2.5-14B and 93.3 percent with Qwen2.5-72B, while consuming roughly one-tenth the tokens of raw trajectory replay.
The interesting finding is not the success-rate bump, it is where the gain comes from: controlled comparisons show verified action normalization, not better retrieval or formatting, does most of the work in making old experience usable again. That matters because most memory research for AI agents has focused on how to store and fetch experience, not on whether a retrieved memory is actually safe to execute as-is.
A successful run from last week can still be the wrong move today, and the fix here is less about smarter search than about editing an agent's memory before it acts on it.