菜单
Create a universal memory infrastructure enabling agents across different frameworks to share experience trajectories without retraining. Improve agent performance by retrieving workflows from related domains and applying diagnostic fixes.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill cross-domain-agent-knowledge命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | cross-domain-agent-knowledge |
| title | Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2507.06229 |
| keywords | ["Agent Memory","Knowledge Transfer","Cross-Domain Learning","Agent Frameworks","Experience Sharing"] |
| description | Create a universal memory infrastructure enabling agents across different frameworks to share experience trajectories without retraining. Improve agent performance by retrieving workflows from related domains and applying diagnostic fixes. |
Agent frameworks like smolagents and OpenHands operate in isolation, causing each agent to rediscover solutions and repeat mistakes independently. Agent KB establishes a shared knowledge repository that allows agents across heterogeneous frameworks to benefit from experience accumulated in other domains. By implementing hybrid retrieval (for planning seeds and feedback fixes) with a safeguard mechanism that prevents harmful knowledge transfer, the system achieves substantial improvements without requiring framework modifications or model retraining.
The core problem is knowledge silos: agents accumulate valuable trajectories but cannot share them across framework boundaries, creating duplicated effort and preventing collective intelligence from emerging.
Agent KB operates on three key principles:
This architecture enables seamless knowledge sharing without requiring agents to be aware of framework differences.
Set up the knowledge base infrastructure to aggregate trajectories from multiple sources:
from agent_kb.knowledge_base import KnowledgeBase
from agent_kb.standardizers import TrajectoryStandardizer
# Initialize universal knowledge repository
kb = KnowledgeBase(storage=)
standardizer = TrajectoryStandardizer()
trajectories_smolagents = load_smolagents_trajectories()
trajectories_openhands = load_openhands_trajectories()
trajectory trajectories_smolagents + trajectories_openhands:
standardized = standardizer.standardize(
trajectory=trajectory,
framework=
)
kb.add_trajectory(
trajectory=standardized,
metadata={
: trajectory.get(),
: trajectory.get(),
: trajectory.get(),
: trajectory.get()
}
)
()
Implement planning-stage retrieval that seeds new agents with relevant workflows:
from agent_kb.retrievers import PlanningRetriever
retriever = PlanningRetriever(kb=kb)
def solve_with_planning_seeds(task_description, agent):
"""Augment agent planning with retrieved seed workflows."""
# Retrieve high-level workflows from successful trajectories
seed_workflows = retriever.retrieve_planning_seeds(
query=task_description,
k=3, # Top 3 relevant workflows
filter_success=True # Only successful trajectories
)
# Extract workflow skeleton (task sequence, not details)
workflows = [
{
"task_sequence": extract_task_steps(traj),
"estimated_success_rate": traj["metadata"]["success_rate"],
"domain": traj["metadata"]["domain"]
}
for traj in seed_workflows
]
# Provide as planning context to agent
agent.set_planning_seeds(workflows)
# Agent uses seeds to guide initial action sequence
result = agent.solve(task_description)
return result
Implement feedback-stage retrieval that suggests fixes after agent failures:
from agent_kb.retrievers import FeedbackRetriever
from agent_kb.diagnostics import DiagnosticExtractor
feedback_retriever = FeedbackRetriever(kb=kb)
extractor = DiagnosticExtractor()
def recover_from_failure(failed_trajectory, agent, kb):
"""Use KB diagnostics to suggest recovery actions."""
# Identify failure pattern in current trajectory
failure_pattern = extractor.extract_failure_pattern(
trajectory=failed_trajectory,
agent_state=agent.current_state
)
# Find similar failures in KB and their resolutions
diagnostic_fixes = feedback_retriever.retrieve_fixes(
failure_pattern=failure_pattern,
k=5 # Top 5 similar failure modes
)
# Rank fixes by applicability to current context
applicable_fixes = [
fix for fix in diagnostic_fixes
if is_applicable(fix, agent.current_state)
]
if not applicable_fixes:
return None # No relevant fixes found
# Format as suggestions to agent
recovery_suggestions = [
{
"action": fix["recovery_action"],
"reasoning": fix["why_worked"],
"confidence": fix["applicability_score"]
}
for fix in applicable_fixes[:3]
]
return recovery_suggestions
Implement the disagreement gate to prevent negative transfer:
from agent_kb.safety import DisagreementGate
gate = DisagreementGate()
def apply_retrieved_knowledge(retrieved_knowledge, agent_state, current_task):
"""Apply retrieved knowledge only if it aligns with current reasoning."""
# Check for disagreement between retrieved knowledge and agent state
disagreement_score = gate.evaluate(
retrieved_knowledge=retrieved_knowledge,
agent_reasoning=agent_state.get("reasoning"),
task=current_task
)
# If disagreement is high, suppress the retrieved knowledge
if disagreement_score > gate.threshold:
print(f"Suppressing retrieved knowledge (disagreement: {disagreement_score:.2f})")
return None # Don't apply conflicting knowledge
# Otherwise, integrate into agent planning
enhanced_plan = integrate_knowledge(
agent_plan=agent_state.get("current_plan"),
retrieved_knowledge=retrieved_knowledge
)
return enhanced_plan
Use Agent KB when:
Avoid Agent KB for:
Standard format for all trajectories (framework-agnostic):
{
"task": "Book a flight from NYC to LA",
"domain": "travel",
"framework": "smolagents",
"success": True,
"steps": [
{
"action": "search_flights",
"input": {"origin": "NYC", "destination": "LA", "date": "2024-03-25"},
"output": "Found 12 flights, cheapest $89",
"tool": "FlightSearch"
},
{
"action": "filter_results",
"input": {"max_price": 150, "preferred_airline": "Delta"},
"output": "3 flights match criteria"
}
],
"metadata": {
"total_steps": 2,
"tools_used": ["FlightSearch", "Filter"],
"success_rate": 0.87,
"timestamp": "2024-03-25T10:00:00Z"
}
}
The paper demonstrates substantial gains by framework:
| Framework | Baseline | With KB | Improvement |
|---|---|---|---|
| smolagents | 55.2% | 73.9% | +18.7 pp |
| OpenHands | 24.3% | 28.3% | +4.0 pp |
| Custom agent | Variable | +12-15 pp | Domain dependent |
Larger improvements occur when source and target domains are closely related.
| Stage | Strategy | When to Use |
|---|---|---|
| Planning | Seed workflows | Early in task, need high-level guidance |
| Feedback | Diagnostic fixes | After agent failure, need recovery suggestions |
| Adaptation | Full trajectory | When source task is very similar to target |
For KB scaling beyond 100k trajectories:
"Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving" - arXiv:2507.06229