agentic-science-cognitive-accumulation

name	agentic-science-cognitive-accumulation
title	Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2601.10402
keywords	["agentic-science","long-horizon","context-management","knowledge-distillation","ML-engineering"]
description	Enables agents to maintain strategic coherence over extended experimental cycles through hierarchical cognitive caching that distills execution traces into stable knowledge, achieving 56.44% on MLE-Bench within 24-hour budgets.

Overview

Implement a hierarchical cognitive caching system for autonomous agents conducting multi-day ML engineering experiments. Rather than maintaining static context windows, the system dynamically distills execution traces into reusable knowledge representations, allowing agents to decouple immediate execution from long-term experimental strategy.

When to Use

• For multi-day autonomous research or engineering projects requiring hundreds of experimental steps
• When agents need to explore high-dimensional problem spaces beyond human precedent
• For ML hyperparameter tuning, architecture search, or scientific discovery tasks
• When you need agents to learn from prior experimental failures and optimize future attempts

When NOT to Use

• For short-horizon tasks (single-session experiments under 1 hour)
• When all relevant context fits in static context windows
• For real-time systems where knowledge consolidation adds unacceptable latency
• For tasks with simple, deterministic experimental spaces

Key Technical Components

Hierarchical Cognitive Caching (HCC)

Implement multi-tier knowledge distillation that converts verbose execution traces into compressed representations.

# Three-tier hierarchy for knowledge consolidation
class HierarchicalCache:
    def __init__(self):
        self.immediate_context = {}  # Current task execution
        self.session_knowledge = {}    # Session-level insights
        self.cross_task_insights = {}  # Lessons across experiments

    def consolidate_traces(self, execution_trace, level="session"):
        """Distill trace into stable knowledge at specified level"""
        if level == "session":
            # Abstract specific values to principles
            pattern = self.extract_principles(execution_trace)
            .session_knowledge.update(pattern)
         level == :
            
            insight = .abstract_to_meta_strategy(execution_trace)
            .cross_task_insights.update(insight)