메뉴
Enable long-horizon agents to manage finite context by separating working memory from persistent storage. Use indexed summaries with pointers to archived evidence, treating memory operations as first-class agent actions learned via RL.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill memory-indexed-experience-scaling명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
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.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | memory-indexed-experience-scaling |
| title | Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.04257 |
| keywords | ["Long-Horizon Reasoning","Memory Management","Context Compression","Agent Scaling","Reinforcement Learning"] |
| description | Enable long-horizon agents to manage finite context by separating working memory from persistent storage. Use indexed summaries with pointers to archived evidence, treating memory operations as first-class agent actions learned via RL. |
Long-horizon multi-step reasoning tasks overwhelm finite context windows: agents either repeat full conversation history (inefficient) or use lossy summaries (information loss). Memex introduces a memory architecture where agents explicitly manage both working context and long-term storage through learned operations. The system maintains compact in-context indexed summaries with pointers to externally archived artifacts, enabling full-fidelity evidence retrieval on demand.
The core innovation treats memory as a first-class agent capability: read operations dereference indices to retrieve exact past information; write operations compress state into indexed summaries. This mirrors how humans manage complex projects with external notes and references while keeping working memory focused.
Memex decomposes memory into two complementary systems:
Agents learn to:
This separation enables agents to maintain reasoning coherence over thousands of tokens while staying within practical context limits.
Step 1: Design experience store and indexing scheme
Create a persistent storage system for archived experiences with stable access keys.
class ExperienceStore:
"""External memory store for agent artifacts."""
():
.store = {}
.index_counter =
.max_size = max_size
():
(.store) >= .max_size:
oldest_idx = (.store.keys(),
key= k: .store[k][])
.store[oldest_idx]
idx = .index_counter
.store[idx] = {
: artifact,
: artifact_type,
: time.time(),
: time.time(),
:
}
.index_counter +=
idx
():
idx .store:
entry = .store[idx]
entry[] = time.time()
entry[] +=
entry[]
():
(.store.keys())
Step 2: Create working context with indexed summaries
Design a compact in-context representation that embeds indices without storing full artifacts.
def create_indexed_summary(current_trajectory, experience_indices,
summary_max_tokens=200):
"""
Create compact working context with embedded index references.
Format:
- Current step context (80-100 tokens)
- Index references to archived experiences (20-40 tokens)
- Metadata about available archived artifacts (remaining tokens)
"""
# Part 1: Recent context (last 2-3 steps)
recent_context = format_recent_trajectory(current_trajectory, depth=3)
# Part 2: Archived experience index pointers
index_pointers = []
for idx in experience_indices[-5:]: # Reference last 5 archived experiences
summary = f"[exp_{idx}] " # Compact reference
index_pointers.append(summary)
# Part 3: Metadata about available memories
available_memories = f"Available: {len(experience_indices)} archived experiences"
# Combine into compact summary
working_context = f"""
Current Context:
{recent_context}
Available Memories:
{' '.join(index_pointers)}
{available_memories}
Agent Action Space: [act, read_exp_IDX, write_summary]
"""
return working_context.strip()
def format_recent_trajectory(trajectory, depth=3):
"""Format last `depth` steps of trajectory for context."""
recent = trajectory[-depth:]
formatted = []
for step in recent:
formatted.append(f"Step: {step['action']}, Result: {step['result'][:50]}")
return '\n'.join(formatted)
Step 3: Implement read/write agent actions
Extend agent action space to include memory operations as learned behaviors.
class MemoryAugmentedAgent:
"""LLM agent with learned memory operations."""
def __init__(self, llm_model, experience_store, working_context_size=200):
self.llm = llm_model
self.store = experience_store
self.working_context_size = working_context_size
def step(self, task_prompt, working_memory, trajectory):
"""
Execute agent step with optional memory operations.
Returns: (action, memory_operation)
"""
# Construct full prompt with working memory
indexed_summary = create_indexed_summary(trajectory, self.store.list_indices())
full_prompt = f"""
Task: {task_prompt}
Working Memory:
{indexed_summary}
Your action (format: ACTION:value or READ:exp_IDX or WRITE:summary_text):
"""
# Forward pass through LLM
response = self.llm.generate(full_prompt, max_tokens=100)
# Parse response for action type
if response.startswith('READ:'):
# Memory read operation
exp_idx = int(response.split(':')[1])
artifact = self.store.read(exp_idx)
return ('read', artifact, exp_idx)
elif response.startswith('WRITE:'):
# Memory write operation
summary = response.split(':', 1)[1]
idx = self.store.write(summary, artifact_type='agent_summary')
return ('write', idx)
else:
# Regular action
return ('act', response)
def compress_for_storage(self, trajectory_segment):
"""
Compress trajectory segment into archived summary.
Executed when agent issues WRITE action.
"""
# Summarize key decisions and outcomes
summary = f"""
Trajectory Summary:
- Steps: {len(trajectory_segment)}
- Key decisions: {[s['action'] for s in trajectory_segment[:3]]}
- Final outcome: {trajectory_segment[-1]['result']}
"""
return summary
Step 4: Learn memory operations via reinforcement learning
Train agent to optimize when and what to read/write using RL.
def compute_memory_cost(operation_type, access_count=0):
"""
Compute cost of memory operation in tokens.
- read: 1-2 tokens (index) + retrieved artifact tokens
- write: summary length + index (2-3 tokens)
"""
if operation_type == 'read':
# Cost: index overhead + artifact retrieval
return 2 + access_count # Penalize repeated reads of same info
elif operation_type == 'write':
# Cost: summary compression + indexing overhead
return 3 + 0 # Fixed overhead
else:
return 0 # No memory cost for regular actions
def compute_memory_reward(trajectory, task_success):
"""
Reward successful task completion while penalizing memory operations.
reward = task_success_bonus - memory_operation_cost
"""
success_bonus = 10.0 if task_success else -1.0
# Count memory operations in trajectory
total_memory_cost = 0
for step in trajectory:
if step['action_type'] in ['read', 'write']:
total_memory_cost += compute_memory_cost(step['action_type'])
# Memory efficiency reward
memory_penalty = 0.1 * total_memory_cost
return success_bonus - memory_penalty
def memory_rl_update(agent, batch_trajectories, learning_rate=0.001):
"""
RL update to optimize memory operations.
Uses policy gradient with memory cost penalties.
"""
total_pg_loss = 0.0
for trajectory in batch_trajectories:
task_success = trajectory[-1]['task_complete']
reward = compute_memory_reward(trajectory, task_success)
# Policy gradient for each step
for t, step in enumerate(trajectory):
# Log probability of chosen action
logprob = agent.llm.compute_logprob(
step['prompt'],
step['action']
)
# PG loss: maximize log-prob of low-cost, high-reward actions
step_memory_cost = compute_memory_cost(step['action_type'])
adjusted_reward = reward - step_memory_cost
pg_loss = -logprob * adjusted_reward
total_pg_loss += pg_loss
# Optimize
agent.llm.backward(total_pg_loss / len(batch_trajectories))
agent.llm.optimizer.step(learning_rate)
return (total_pg_loss / len(batch_trajectories)).item()
Step 5: Scaling evaluation on long-horizon tasks
Benchmark memory-augmented agents on tasks requiring hundreds of steps.
def evaluate_scaling(agent, task_sequence, max_steps=500):
"""
Evaluate agent on long-horizon task.
Compare: working context size, success rate, memory operations.
"""
trajectory = []
working_memory = []
context_size = 0
success = False
for step_idx in range(max_steps):
# Agent step
action, memory_op = agent.step(
task_sequence[step_idx],
working_memory,
trajectory
)
trajectory.append({
'step': step_idx,
'action': action,
'memory_op': memory_op,
'context_size': context_size
})
# Update working memory
if memory_op[0] == 'write':
working_memory.append(memory_op[1])
context_size += len(str(memory_op[1]))
# Check task completion
if agent.check_success(action):
success = True
break
return {
'success': success,
'steps_taken': len(trajectory),
'peak_context_size': max(s['context_size'] for s in trajectory),
'memory_ops': len([t for t in trajectory if t['memory_op'] is not None]),
'trajectory': trajectory
}
Hyperparameter Selection:
When to Use:
When NOT to Use:
Common Pitfalls: