Behavior cloning (BC) in non‑Markovian settings requires reasoning over long‑horizon context. Existing policies rely on recurrent or attention mechanisms, yet recurrent models suffer hidden‑state collapse and gradient instability during back‑propagation through time, while attention models are fundamentally limited by context length.
We introduce Keyframe Mnemonics, a self‑supervised approach that discovers information‑critical observations (mnemonics) by learning an objective from randomly sampled past observations and using it as a reward for keyframe selection. A BC policy is then trained conditioned on the discovered keyframes to model the action distribution.
Under certain task‑structure assumptions, the formulation guarantees context retention over an infinite horizon while keeping only a small set of decision‑relevant keyframes in the policy’s working memory.
Evaluations on synthetic memory domains show mnemonic‑conditioned BC policies achieve $100\%$ success rate (SR) and generalize to horizons orders of magnitude beyond training without degradation. On a memory‑intensive robot manipulation benchmark, the method yields a $13.9\%$ average absolute SR improvement over the strongest baseline across 23 tasks and retains $80\%$ SR at $20\times$ longer horizons on a real robot. Code and videos are available at https://keyframe-mnemonics.github.io.
Review: By leveraging self‑supervised keyframe selection, this work overcomes the inherent limits of recurrent and attention‑based models, offering a scalable and efficient memory mechanism for long‑horizon behavior cloning.