An AI agent can remember facts, past experiences, or how-to routines — what does a saved 'recipe' add that the others don't?
How do workflow and function memories contribute differently in agent learning?
This explores what a workflow memory (a reusable how-to routine) adds to an agent compared with the other jobs memory does, which the corpus calls 'functions': holding facts, recording experience, and tracking the current task.
This explores what a workflow memory (a reusable how-to routine) adds to an agent compared with the other jobs memory does, which the corpus calls 'functions': holding facts, recording experience, and tracking the current task. The corpus has no note that sets the two side by side, so what follows is assembled from separate findings. One survey splits agent memory by function into factual (what's true), experiential (what happened and what worked), and working (what's in play right now), and argues this split is more useful than short-term versus long-term Can three axes replace the short-term long-term memory split?. On that map, workflow memory is a specific kind of experiential memory: it stores how to do things, not what is true.
What sets workflow memory apart is its granularity and how it generalizes. Agent Workflow Memory pulls out sub-task routines, which are smaller than whole tasks. It strips the example-specific values so a routine like 'search, filter, then open the result' can be reused on a different site, and it stacks routines into bigger ones Can agents learn reusable sub-task routines from past experience?. It gained 24.6% on Mind2Web and 51.1% on WebArena, and the gains were larger when the test tasks looked less like the training tasks. That pattern suggests the routine, not the memorized example, is what transfers. Compare Reflexion, another experiential memory, which stores a verbal self-diagnosis of one specific failure and helps the next attempt at that same task Can agents learn from failure without updating their weights?. A reflection says what went wrong here. A workflow says what usually works.
The other functions do different work. AgentFly splits memory into case, subtask, and tool modules and improves the agent purely through reading and writing them, reaching 87.88% on GAIA with no weight updates Can agents learn continuously from experience without updating weights?. Working memory is where things break, and better routines can't fix that. One paper finds long-workflow failures come from missing gates on what gets written and recalled, not from missing knowledge. Its fix is a small committed state, separate from permanent memory, that stops errors and drifting constraints from accumulating Can agents fail from weak memory control rather than missing knowledge?. An agent can hold a good routine and still derail halfway through if its current-state memory is leaky.
Skill libraries are the closest relative to workflow memory, and they compound in the same way. VOYAGER stores executable skills, builds complex ones from simpler ones, and refines them with environmental feedback, which avoids the forgetting that weight updates cause Can agents learn new skills without forgetting old ones?. A broader framing treats memory (state), skills (procedures), and protocols (interaction rules) as burdens moved out of the model and into the harness Where does agent reliability actually come from?. In that view, workflow memory belongs with skills, and factual and working memory belong with state. They fix different weaknesses: reusable procedure for tasks the agent hasn't seen, and clean state for staying on track.
The corpus doesn't isolate how much each contributes. That would take module-by-module evaluation, which is how one paper argues memory should be measured. It breaks memory into storage, extraction, retrieval, and maintenance so you can see which part failed, where end-to-end task scores can't How should we actually evaluate agent memory systems?. Until someone runs that comparison across function types, the split is a strong hypothesis, not a measured result.
Sources 8 notes
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.
Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
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.
Agent performance degrades in long workflows because transcript replay and retrieval-based memory lack gating mechanisms. A bounded, schema-governed committed state that separates artifact recall from permanent memory write prevents error accumulation and constraint drift.
Show all 8 sources
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.
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.
Decomposing memory into storage, extraction, retrieval, and maintenance stages exposes design trade-offs and failure modes that task-success metrics completely hide. Module-by-module evaluation across 12 systems shows which component actually failed rather than just whether the task succeeded.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Useful Memories Become Faulty When Continuously Updated by LLMs
- Demystifying Agent Skills: Why They Work-Until They Don't
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Are We Ready For An Agent-Native Memory System?
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments