Experience-RAG Skill: Reusable Procedural Guidance

• Experience-RAG Skill is a framework that retrieves and reuses distilled procedural assets from previous interactions to guide inference, moving beyond static document retrieval.
• It employs diverse representations such as skill memories, detection rules, composite tools, and latent embeddings to support task-specific, context-aware decision making.
• Retrieval is conditioned on the agent's task state and governed by mechanisms for routing, maintenance, and evaluation, ensuring reliable and transferable performance across domains.

Searching arXiv for papers on Experience-RAG Skill and closely related frameworks. Experience-RAG Skill denotes a class of retrieval-augmented mechanisms in which an agent retrieves and reuses prior experience as procedural guidance rather than relying only on static documents. Across recent work, the retrieved unit may be a structured skill, a rule, a composite tool, a navigable summary node, a verified trajectory fragment, or a latent experience embedding, but the common objective is consistent: transform prior successful interactions into reusable competence that can be selected, injected, executed, audited, or refined at inference time without necessarily updating model weights (Sun et al., 8 Jun 2026, Zhang et al., 5 May 2026, Feng et al., 31 May 2026).

1. Definition and conceptual scope

Experience-RAG differs from conventional document-centric RAG by retrieving “how to act” in addition to “what to know.” In the medical-agent setting of SkeMex, Experience-RAG means retrieval-augmented generation over distilled agent experience rather than static documents; prior interaction trajectories are converted into structured, reusable skills and retrieved at inference time to condition reasoning and tool use (Sun et al., 8 Jun 2026). In the retrieval-orchestration setting, Experience-RAG is a pluggable layer between the agent and the retriever pool that analyzes the current scene, consults an experience memory, selects an appropriate retrieval strategy, and returns structured evidence (Zhang et al., 5 May 2026). In latent formulations such as ExpWeaver, the same idea is realized without text concatenation: experiences are encoded using the model’s own hidden states, retrieved directly in latent space at each decoding step, and integrated by cross-attention and gated residual mechanisms (Feng et al., 31 May 2026).

The core memory unit varies by system. SkeMex defines each memory as a skill $m=(k,c,u)$ with retrieval key $k$, procedural content $c$, and empirical utility $u$ estimated from interaction outcomes (Sun et al., 8 Jun 2026). MA-IDS stores an experience entry as a vector key plus a metadata payload containing a human-readable detection rule and the labels $(\hat{y},y)$ that triggered rule induction (Islam et al., 7 Apr 2026). HERA stores Profile–Insight–Utility triplets $(c,z,u)$, where $c$ is a query profile, $z$ is a distilled natural-language insight, and $u$ is an empirical success count (Li et al., 1 Apr 2026). Anything2Skill compiles external knowledge into explicit skill contracts with invocation conditions, contraindications, workflow steps, constraints, outputs, supporting evidence, and confidence scores (Pan et al., 8 Jun 2026).

A common misconception is that Experience-RAG is simply long-term memory or raw trace replay. Multiple papers explicitly oppose that interpretation. SkeMex argues that raw trace replay stores long, noisy, instance-bound trajectories that are hard to govern and lack transferability, whereas skill memory abstracts trajectories into compact procedures and governs reuse by measured utility (Sun et al., 8 Jun 2026). MA-IDS likewise externalizes learning into an Experience Library of reusable rules rather than modifying model weights (Islam et al., 7 Apr 2026). Corpus2Skill further reframes retrieval itself as navigation over distilled corpus structure rather than repeated flat search (Sun et al., 16 Apr 2026). This suggests that the defining feature of Experience-RAG is not storage alone, but selective abstraction of experience into reusable, governable, and task-relevant assets.

2. Representational forms of experience

Experience-RAG systems differ primarily in what they regard as a reusable unit of experience.

SkeMex organizes experience as multi-branch skill memory spanning general, task-specific, and action-level branches, with formal partition

$M_t = M_t^{gen} \cup M_t^{task} \cup M_t^{act}.$

Each skill records triggers, steps and tools, expected outcomes or checks, branch labels, clinical category tags, provenance, adoption statistics, and timestamps for memory strength or decay (Sun et al., 8 Jun 2026). D2Skill instead partitions a dynamic skill bank into task skills for high-level guidance and step skills for localized correction, with separate retrieval keys, utilities, and update rules for each granularity (Tu et al., 30 Mar 2026). LEGO represents reusable experience as “composable circuit skills,” standardized as a tuple $k$0 and optionally augmented with an experience store $k$1 containing summaries, symptoms, root causes, strategies, applicability guards, and artifacts (Lou et al., 25 Apr 2026).

Other systems compile different structures. MACRO mines verified execution trajectories for recurring contiguous tool subsequences and registers them as composite tools, expanding the agent’s action space over time (Fan et al., 6 Mar 2026). RAG-Modulo stores interaction-level tuples $k$2, where

$k$3

combining goal, previous action, critic feedback, and current observation to support context-aware step selection in sequential decision making (Jain et al., 2024). ASG-SI promotes successful trajectory fragments into auditable skill-graph nodes with explicit interfaces, replayable provenance, and verifier-backed evidence bundles (Huang et al., 28 Dec 2025). Corpus2Skill compiles enterprise corpora into hierarchical skill directories whose nodes are summaries and navigable subgroup files, making corpus organization itself part of the reusable skill surface (Sun et al., 16 Apr 2026).

A concise way to compare these forms is to distinguish whether the reused object is primarily procedural, diagnostic, strategic, structural, or latent.

Form of experience	Representative system	Retrieved object
Structured procedural skill	SkeMex, Anything2Skill	Skill contract or skill memory
Error-derived rule	MA-IDS	Human-readable detection rule
Dual-granularity guidance	D2Skill	Task skill or step skill
Composite action primitive	MACRO	Registered composite tool
Interaction exemplar	RAG-Modulo	Feasible interaction snippet
Navigable corpus abstraction	Corpus2Skill	Skill file / summary node
Latent embedding	ExpWeaver	Hidden-state experience vector

These representations are not interchangeable. Systems operating in environments with typed tools and explicit workflows tend to favor structured contracts or composite tools (Pan et al., 8 Jun 2026, Fan et al., 6 Mar 2026), while systems centered on generation quality or routing often favor summaries, insights, or latent memories (Li et al., 1 Apr 2026, Feng et al., 31 May 2026). A plausible implication is that the optimal representation depends on whether downstream reuse requires execution, diagnosis, retrieval control, or simple contextual biasing.

3. Retrieval, routing, and inference-time conditioning

A central property of Experience-RAG Skill is that retrieval is conditioned on task state, not merely on lexical similarity. SkeMex retrieves once at episode onset and scores each candidate skill by a value-aware combination of semantic similarity, category-conditioned utility, and decayed memory strength:

$k$4

with default weights $k$5, $k$6, $k$7 and $k$8-$k$9 (Sun et al., 8 Jun 2026). D2Skill uses a two-stage retrieval process in which candidates passing a cosine-similarity filter are reranked by a joint similarity-utility score with an exploration bonus:

$c$0

This retrieval is applied separately for task skills and step skills (Tu et al., 30 Mar 2026).

Experience can also control which retriever to use. Experience-RAG Skill for retrieval orchestration defines scene construction and routing as

$c$1

and its strongest configuration uses a rule-based policy:

$c$2

This mechanism achieves an overall $c$3 of $c$4 on BeIR/nq, BeIR/hotpotqa, and BeIR/scifact, outperforming fixed single-retriever baselines and remaining competitive with Adaptive-RAG-style routing (Zhang et al., 5 May 2026).

Failure-state-aware routing is another retrieval-control variant. Skill-RAG probes hidden states before and after retrieval, diagnoses whether failure is due to surface-form divergence, entangled premises, overly broad scope, or an irreducible case, and then routes among four retrieval skills: query rewriting, question decomposition, evidence focusing, and exit (Wei et al., 17 Apr 2026). On out-of-distribution datasets, it reports particularly strong gains over Probing-RAG, including $c$5 ACC on MuSiQue and $c$6 ACC on 2Wiki (Wei et al., 17 Apr 2026). This supports the paper’s claim that post-retrieval failure is typed rather than monolithic.

Latent retrieval moves conditioning inside the decoder itself. ExpWeaver retrieves top-$c$7 experiences at each decoding step using the current hidden state $c$8 as a latent query,

$c$9

aggregates them with cross-attention, and merges them by a norm-preserving gated residual update (Feng et al., 31 May 2026). Because no text snippets are appended, token usage remains comparable to non-retrieval baselines while text-based retrieval methods require $u$0 to $u$1 times more tokens (Feng et al., 31 May 2026).

4. Writing, valuation, and continual maintenance

Experience-RAG Skill systems generally include a write path that turns new interactions into reusable memory and a governance path that prevents uncontrolled growth or drift.

SkeMex formalizes continual evolution as a closed-loop lifecycle

$u$2

where $u$3 is Writing, $u$4 is Valuation, and $u$5 is Governance (Sun et al., 8 Jun 2026). Trajectories enter a gated buffer according to

$u$6

which favors informative multi-step traces while balancing categories (Sun et al., 8 Jun 2026). Utilities are then updated from adoption credit based on trajectory advantage relative to a per-category EMA baseline, with clipping to $u$7, default $u$8, $u$9, $(\hat{y},y)$0, and EMA $(\hat{y},y)$1 (Sun et al., 8 Jun 2026). Governance merges near-duplicates at similarity $(\hat{y},y)$2, promotes mature skills with $(\hat{y},y)$3 and usage count $(\hat{y},y)$4, deprecates low-utility skills with $(\hat{y},y)$5, and enforces capacities $(\hat{y},y)$6, $(\hat{y},y)$7 per category, and $(\hat{y},y)$8 per tool (Sun et al., 8 Jun 2026).

D2Skill maintains its bank by reflection, online utility updates, and pruning. For each task $(\hat{y},y)$9, it compares baseline and skill-injected rollouts, computes a task-level hindsight signal

$(c,z,u)$0

and updates utilities for task skills and step skills using exponential moving averages (Tu et al., 30 Mar 2026). Reflection is triggered when skill-group performance falls below a threshold, producing one task skill and one step skill from failed and successful trajectories (Tu et al., 30 Mar 2026). Skills are then pruned when pool capacity is exceeded, except for recently created skills protected for $(c,z,u)$1 steps (Tu et al., 30 Mar 2026).

MA-IDS uses misclassifications as its write trigger: the Error Analysis Agent converts an error into a human-readable detection rule stored in a persistent FAISS-based Experience Library for future retrieval (Islam et al., 7 Apr 2026). RAG-Modulo commits only successful, feasible interactions to global memory at task completion, preventing noisy or infeasible steps from polluting the store (Jain et al., 2024). ASG-SI goes further by requiring verifier-backed replay, contract checks, and append-only audit manifests before promotion into the skill graph (Huang et al., 28 Dec 2025). Anything2Skill performs lifecycle actions at registry level, choosing among Create, Strengthen, Revise, Merge, Split, Unchanged, and Discard, while maintaining versioned updates and visible skill-tree projection (Pan et al., 8 Jun 2026).

These mechanisms address another misconception: that more memory is inherently better. Several studies report the opposite. D2Skill shows that dynamic skill maintenance is critical to gains (Tu et al., 30 Mar 2026). SkeMex ablations show substantial drops when maturation, deprecation, or merging are removed (Sun et al., 8 Jun 2026). Corpus2Skill reports that wide-shallow trees with branching ratio $(c,z,u)$2 degrade routing quality relative to narrower hierarchies (Sun et al., 16 Apr 2026). This suggests that Experience-RAG is as much about curation and governance as about accumulation.

5. Empirical patterns across domains

A notable feature of the literature is the breadth of domains in which Experience-RAG Skill has been instantiated.

In medical reasoning, SkeMex reports consistent gains in both offline and online settings across nine medical benchmarks. On a DeepSeek-V3.2 backbone, ReAct improves from $(c,z,u)$3 to $(c,z,u)$4, and on Qwen3.6-Plus from $(c,z,u)$5 to $(c,z,u)$6 (Sun et al., 8 Jun 2026). Under frozen-repository out-of-domain transfer, DeepSeek-V3.2 improves from $(c,z,u)$7 to $(c,z,u)$8, and Qwen3.6-Plus from $(c,z,u)$9 to $c$0 (Sun et al., 8 Jun 2026). Skills learned on one backbone also transfer to Claude Sonnet-4.6 and Qwen3.6-35B-A3B, improving their averages by about $c$1 points over ReAct (Sun et al., 8 Jun 2026).

In medical imaging, MACRO reports that autonomous composite-tool discovery, image-feature memory, and GRPO-like training improve multi-step orchestration across fundus, echocardiography, and radiography tasks. On REFUGE2 it reaches balanced accuracy $c$2 and F1 $c$3, compared with a baseline without memory, composites, or GRPO at BACC $c$4 and F1 $c$5 (Fan et al., 6 Mar 2026). On MITEA, the full system achieves BACC $c$6 and F1 $c$7, up from $c$8 and $c$9 in the baseline (Fan et al., 6 Mar 2026).

In intrusion detection, MA-IDS uses an Experience Library of rules extracted from misclassifications and reports Macro F1 of $z$0 on NF-BoT-IoT and $z$1 on NF-ToN-IoT, compared with zero-shot GPT-4o baselines of $z$2 and $z$3 (Islam et al., 7 Apr 2026). “Library Only” already captures most of the lift, indicating that externalized error experience is the dominant factor (Islam et al., 7 Apr 2026).

In retrieval orchestration, Experience-RAG Skill reaches Recall@10 $z$4, MRR@10 $z$5, and $z$6 $z$7 on mixed BeIR workloads, outperforming fixed dense, BM25, and hybrid_rrf baselines on overall workload quality (Zhang et al., 5 May 2026). In enterprise QA, Corpus2Skill reports on WixQA that navigable corpus skills achieve F1 $z$8, Factuality $z$9, and Context Recall $u$0, outperforming dense retrieval, RAPTOR, and agentic RAG baselines across all quality metrics (Sun et al., 16 Apr 2026).

In sequential decision making, RAG-Modulo on BabyAI-BossLevel improves success rate from $u$1 for LLM-Planner to $u$2, reduces average infeasible actions from $u$3 to $u$4, and shortens average episode length from $u$5 to $u$6 (Jain et al., 2024). D2Skill reports 10–20 point gains over skill-free baselines on ALFWorld and WebShop, including ALFWorld overall success $u$7 versus GRPO $u$8 on Qwen2.5-7B-Instruct (Tu et al., 30 Mar 2026).

In compiled external knowledge, Anything2Skill combined with RAG achieves success rates of $u$9 on qsv and $M_t = M_t^{gen} \cup M_t^{task} \cup M_t^{act}.$0 on GitHub-CLI, substantially outperforming both the Base Agent and RAG-only settings (Pan et al., 8 Jun 2026). This is one of the clearest demonstrations that declarative passage retrieval and compiled procedural skill retrieval are complementary rather than competing modes.

6. Design tensions, limitations, and outlook

Several recurring tensions define the current state of Experience-RAG Skill.

The first is abstraction versus fidelity. Raw trajectories preserve detail but are noisy and hard to govern; distilled skills or rules are compact and reusable but can omit edge conditions (Sun et al., 8 Jun 2026, Islam et al., 7 Apr 2026). Anything2Skill addresses this by retaining strict evidence windows and supporting evidence sets for every compiled skill (Pan et al., 8 Jun 2026), while ASG-SI insists on replayable provenance and reconstructible rewards (Huang et al., 28 Dec 2025). This suggests that future systems will likely continue to separate a compact reusable artifact from a richer audit trail.

The second is adaptivity versus stability. Systems such as SkeMex, D2Skill, HERA, and MACRO continuously expand or revise their memory without weight updates or with lightweight policy optimization (Sun et al., 8 Jun 2026, Tu et al., 30 Mar 2026, Li et al., 1 Apr 2026, Fan et al., 6 Mar 2026). Yet all four identify drift, noisy utilities, or misattribution as failure modes. SkeMex explicitly notes utility noise for rare categories, overfitting to overrepresented patterns, runtime overhead, and privacy risks if real data are used (Sun et al., 8 Jun 2026). HERA notes attribution noise and sensitivity to memory hygiene (Li et al., 1 Apr 2026). D2Skill highlights utility drift and skill misapplication as practical risks (Tu et al., 30 Mar 2026).

The third is textual explicitness versus latent efficiency. Most systems retrieve explicit text or structured contracts, preserving interpretability and easy governance (Sun et al., 8 Jun 2026, Pan et al., 8 Jun 2026, Islam et al., 7 Apr 2026). ExpWeaver instead retrieves in latent space and reports state-of-the-art performance on 12 of 13 tasks, more than $M_t = M_t^{gen} \cup M_t^{task} \cup M_t^{act}.$1 improvement over the strongest baseline on average, superior zero-shot and few-shot transfer, and token efficiency comparable to non-retrieval baselines (Feng et al., 31 May 2026). A plausible implication is that the field may bifurcate into interpretable explicit-memory systems for high-stakes domains and latent end-to-end systems for efficiency-sensitive settings.

The fourth is retrieval versus navigation. Corpus2Skill argues that dense retrieval leaves the model as a passive consumer of results, unable to reason about unexplored branches, whereas a visible hierarchical skill directory enables backtracking and cross-branch evidence synthesis (Sun et al., 16 Apr 2026). This claim is consistent with broader Experience-RAG trends: reusable experience need not always be a snippet or a rule; it may be a navigable organizational structure.

Across the literature, the most stable conclusion is not that any single architecture is dominant, but that retrieval-augmented agents improve when experience is transformed into reusable procedural assets and governed over time. Whether those assets are skills, rules, insights, composite tools, routed retrieval strategies, or latent embeddings, the field converges on a shared principle: experience becomes useful when it is abstracted enough to transfer, structured enough to retrieve, and controlled enough to remain reliable (Sun et al., 8 Jun 2026, Zhang et al., 5 May 2026, Pan et al., 8 Jun 2026).