Can an AI searcher remember which move it made, like widening or backtracking, and mix those lessons into new strategies?

This explores whether memory tied to specific search moves (the 'operators' an agent applies, like expanding, refining or backtracking) helps a search agent recombine what it has learned instead of replaying one fixed strategy. The corpus has no note that tests operator-conditioned memory head-to-head, so there is no direct yes or no. Three neighboring threads bear on the underlying problem, and together they make the idea plausible.

The closest working example of composition through memory is VOYAGER. It stores behaviors as executable skills in an embedding-indexed library and builds complex skills from simpler ones, so the agent keeps learning without the forgetting that weight updates cause (Can agents learn new skills without forgetting old ones?). That is memory organized around what a behavior does and retrieved by the situation, which is the core of your question. A parallel appears inside networks. Looped models re-apply the same layers repeatedly and gain compositional generalization that adding parameters doesn't buy (Can models learn by looping instead of growing larger?). In both cases, reusable operations applied in sequence beat a single monolithic policy.

What goes into memory also matters. SkillRL treats successful episodes as concrete demonstrations and failures as abstracted lessons. This beats uniform consolidation and uses much less context (Should successful and failed episodes be processed differently?). Reflexion keeps verbal self-diagnoses uncompressed and triggers them with unambiguous success/failure signals, which stops the agent from rationalizing (Can agents learn from failure without updating their weights?). Both results say that memory conditioned on something (outcome type, feedback type) beats a flat pile of experience. Conditioning on the operator would add one more axis of the same kind. The corpus doesn't show that this axis helps, only that other conditioning axes do.

Search agents have a particular reason to want this. RL training squeezes their exploration diversity through the same entropy collapse seen in reasoning, and policies settle on narrow reward-maximizing routes, while SFT on diverse demonstrations preserves breadth (Does reinforcement learning squeeze exploration diversity in search agents?). Composition needs a varied set of behaviors to combine, so a store of operators outside the weights could keep alive what RL would otherwise prune. A two-phase picture of RL points the same way. Training first masters execution, then hits a bottleneck in strategic planning, meaning the choice of what to do next (Does RL training follow a predictable two-phase learning sequence?). Choosing which operator to apply is that planning step, so operator-keyed memory would target the phase where models get stuck. That link is an inference from these notes, not something any of them measured.

There is a broader reason to expect operators to be worth storing. Reasoning generalizes best from broad, transferable procedural knowledge rather than from memorized facts (Does procedural knowledge drive reasoning more than factual retrieval?). Operators are procedures, and that is the kind of knowledge that transfers across problems.

The open gap is a controlled comparison of operator-keyed memory against an unkeyed skill library on search tasks. The notes above make that a well-motivated experiment, but they don't yet show its result.