What distinguishes working memory from strategic memory in agent task execution?
This reads the question as asking how the moment-to-moment state an agent holds while doing a task differs from the higher-level planning and lessons it carries across tasks — and the corpus suggests the line isn't a storage location but a difference in what the memory is *for*.
This explores the split between the transient state an agent uses to execute the step in front of it (working memory) and the accumulated planning knowledge it draws on to decide what to do (strategic memory) — and the most striking thing in the corpus is that this distinction shows up as two separate *learning phases*, not just two storage bins. Across eight models, RL training reliably moves through a first phase where getting the execution right is the bottleneck, followed by a second phase where strategic planning becomes the thing that's hard — and you can watch it happen, because the entropy on planning tokens rises while execution entropy settles down Does RL training follow a predictable two-phase learning sequence?. Working memory is what stabilizes first; strategic memory is what's still being figured out late.
The same asymmetry appears in how agents are taught to *store* the two. SkillRL treats successful episodes as concrete demonstrations you can replay verbatim, and failed episodes as abstracted lessons — exactly the working-vs-strategic divide, where the procedural trace gets kept literally and the strategic insight gets compressed into a rule Should successful and failed episodes be processed differently?. Folding everything through the same consolidation pipeline degrades performance; the two kinds of memory want different treatment because they answer different questions ("how did I do this" vs "what should I do differently").
If you want the cleanest structural map, RAISE decomposes agent working memory into four components on two axes: dialogue-level (the whole conversation, a scratchpad) versus turn-level (examples, the current task trajectory) — and notes that each granularity has its own failure modes and update rules How should agent memory split across time scales?. A broader 2025 survey pushes back on naming any of this "short-term" vs "long-term" at all, arguing the phenomena are better described by *function* — factual, experiential, working — with the temporal feel emerging from how memory forms and gets retrieved rather than from a hard architectural wall Can three axes replace the short-term long-term memory split?. On that view, "strategic memory" is really the experiential/planning function, and "working memory" the active scratchpad function.
The lateral payoff: where strategic knowledge *lives* turns out to depend on the domain. Workflow-level memory wins in routine-rich tasks, causal-rule memory in environment-rich ones, and fine-grained state-action memory in spatially-rich web tasks — the right level of abstraction for your strategic layer is conditional on whether task variance comes from arguments, causal structure, or UI state Does agent memory work better at one level of abstraction?. And the boundary can blur: VOYAGER's externalized skill library is arguably strategic memory made executable — composable procedures that compound over time without the catastrophic forgetting of weight updates Can agents learn new skills without forgetting old ones?, while AgentFly shows an agent can improve its whole policy through memory operations alone, no parameter changes Can agents learn continuously from experience without updating weights?.
The thing worth taking away: the most reliable agents don't hold this distinction in the model's weights at all — they externalize state, skills, and protocols into a harness layer so the model isn't re-solving the same memory problems every turn Where does agent reliability actually come from?. Working vs strategic memory, on that reading, is less a property of cognition than a design decision about which burdens you push outside the model.
Sources 8 notes
Across eight models, RL training consistently shows a first phase where execution correctness drives learning, followed by a second phase where strategic planning becomes the bottleneck. Planning token entropy increases while execution entropy stabilizes, and concentration of optimization on planning tokens yields significant performance gains.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
RAISE shows that agent memory consists of four components organized by two design axes: dialogue-level (conversation history, scratchpad) versus turn-level (examples, task trajectory). This granularity distinction predicts different failure modes and update policies for each component.
A 2025 survey reframes agent memory along forms (token/parametric/latent), functions (factual/experiential/working), and dynamics (formation/evolution/retrieval), showing that short/long-term phenomena emerge from temporal patterns rather than architectural separation. This enables precise system comparison and replaces vague implementation-based claims.
Workflow-level memory wins in routine-rich domains, causal-rule memory in environment-rich domains, and state-action memory in spatially-rich web tasks. The optimal abstraction depends on whether task variance comes from arguments, causal structure, or fine-grained UI state.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.