볼트 홈: README · 조사 방법: 리서치 로그
참고문헌 (Annotated Bibliography) — 마스터 인덱스
이 볼트가 인용한 1차·2차 출처를 분류별로 정리한다. 이 파일은 마스터 인덱스로, 핵심 출처는 여기에 직접 싣고 2026 보강분은 카테고리별 상세 파일로 분리해 링크했다. arXiv 논문은 ID와 제목을 arXiv 공식 API(export.arxiv.org/api)로 대조해 실재를 확인했으며, 기존 74편에 2026 보강분 22편을 더해 누적 96편이다. 블로그와 공식 문서는 각 챕터를 쓰면서 원문을 직접 확인했다.
이 폴더 구성 (_references/)
- sources — (이 파일) 마스터 인덱스. 핵심 1차 출처와 기존 arXiv 74편, 공식 문서·블로그.
- 전체 논문 인덱스 — arXiv 96편 전체를 챕터별 표(ID→arxiv.org 링크, 인용 노트 백링크)로 모은 한 페이지.
- papers-2026-신규 — 2026 보강분 신규 arXiv 22편. 챕터별 표에 ID·제목·한 줄 기여·연결 노트를 담았다.
- articles-2026 — 2026 산업 블로그·엔지니어링 아티클 6편.
- tools-2026 — 2026 Claude/Codex 도구 신기능. 직접 확인한 것과 보도로만 접한 것을 나눠 정리했다.
- 리서치 로그 — 조사 방법, 검증 기록, 미해결 질문.
0. 핵심 1차 — 실패모드 분류 · 정의의 출발점
- Drew Breunig, “How Long Contexts Fail” (2025-06-22) — https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html · 이 볼트가 쓰는 4대 실패모드(오염·산만·혼란·충돌) 분류의 출발점. 01 오염~04 충돌의 토대다.
- Drew Breunig, “How to Fix Your Context” (2025-06-26) — https://www.dbreunig.com/2025/06/26/how-to-fix-your-context.html · 처방 6종(RAG·Tool Loadout·Quarantine·Pruning·Summarization·Offloading). 06 해결전략.
- Chroma Research, “Context Rot: How Increasing Input Tokens Impacts LLM Performance” (Hong·Troynikov·Huber, 2025-07-14) — https://research.trychroma.com/context-rot · 18개 모델에서 성능이 고르지 않게 떨어진다는 실증. 05 부패의 핵심 근거.
- Anthropic, “Effective context engineering for AI agents” (2025) — https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents · 컨텍스트 엔지니어링의 정의와 실천. 00 개요.
- LangChain, “Context Engineering for Agents” (2025-07) — https://www.langchain.com/blog/context-engineering-for-agents · Write/Select/Compress/Isolate 4전략을 처음 제시했다. 06 해결전략.
- Karpathy / Tobi Lütke (2025-06, X) — “context engineering” 용어를 퍼뜨린 글. https://x.com/karpathy/status/1937902205765607626 · https://x.com/tobi/status/1935533422589399127
- Simon Willison, “Context engineering” (2025-06-27) — https://simonwillison.net/2025/Jun/27/context-engineering/ · ‘lethal trifecta’ 개념. 01 오염.
- Philipp Schmid, “Context Engineering” (2025-06-30) — https://www.philschmid.de/context-engineering
- Lance Martin, “Context Engineering” (2025-06-23) — https://rlancemartin.github.io/2025/06/23/context_engineering/
1. 학술 논문 (arXiv)
기초 — Transformer · 컨텍스트 엔지니어링 (00 개요)
- arXiv:1706.03762 — Attention Is All You Need
- arXiv:2209.11895 — In-context Learning and Induction Heads
- arXiv:2507.13334 — A Survey of Context Engineering for Large Language Models
- arXiv:2510.26493 — Context Engineering 2.0: The Context of Context Engineering
장문맥 저하 · 위치편향 · 부패 (02 산만 · 05 부패)
- arXiv:2307.03172 — Lost in the Middle: How Language Models Use Long Contexts (Liu et al., TACL 2024)
- arXiv:2311.09198 — Never Lost in the Middle: Position-Agnostic Decompositional Training
- arXiv:2403.04797 — Found in the Middle: Plug-and-Play Positional Encoding
- arXiv:2603.10123 — Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias
- arXiv:2410.18745 — Why Does the Effective Context Length of LLMs Fall Short?
- arXiv:2502.05167 — NoLiMa: Long-Context Evaluation Beyond Literal Matching
- arXiv:2510.05381 — Context Length Alone Hurts LLM Performance Despite Perfect Retrieval
- arXiv:2506.11440 — AbsenceBench: Language Models Can’t Tell What’s Missing
- arXiv:2302.00093 — Large Language Models Can Be Easily Distracted by Irrelevant Context
장문맥 벤치마크 (05 부패)
- arXiv:2404.06654 — RULER: What’s the Real Context Size of Your Long-Context Language Models?
- arXiv:2410.02694 — HELMET: How to Evaluate Long-Context Language Models Effectively and Thoroughly
- arXiv:2410.10813 — LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
- arXiv:2411.03538 — Long Context RAG Performance of Large Language Models
오염 · 공격 (01 오염)
- arXiv:2507.06261 — Gemini 2.5 (기술보고서; ‘context poisoning’ 포켓몬 사례 출처)
- arXiv:2402.07867 — PoisonedRAG: Knowledge Corruption Attacks to RAG
- arXiv:2407.12784 — AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases
- arXiv:2512.16962 — MemoryGraft: Persistent Compromise of LLM Agents via Poisoned Experience Retrieval
- arXiv:2601.05504 — Memory Poisoning Attack and Defense on Memory Based LLM-Agents
- arXiv:2506.23260 — From Prompt Injections to Protocol Exploits: Threats in LLM-Powered AI Agents Workflows
- arXiv:2503.18813 — Defeating Prompt Injections by Design (CaMeL)
- arXiv:2601.02371 — Permission Manifests for Web Agents
혼란 · 툴 과부하 (03 혼란)
- arXiv:2411.15399 — Less is More: Optimizing Function Calling for LLM Execution on Edge Devices
- arXiv:2505.03275 — RAG-MCP: Mitigating Prompt Bloat in LLM Tool Selection via RAG
- arXiv:2505.06416 — ScaleMCP: Dynamic and Auto-Synchronizing MCP Tools for LLM Agents
- arXiv:2605.24660 — How Many Tools Should an LLM Agent See? A Chance-Corrected Answer
- arXiv:2606.06284 — ToolChoiceConfusion: Causal Minimal Tool Filtering for Reliable LLM Agents
- arXiv:2602.20426 — Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use
- arXiv:2601.05214 — Internal Representations as Indicators of Hallucinations in Agent Tool Selection
- arXiv:2606.10209 — Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents
충돌 · 멀티턴 · 앵커링 · 아첨 (04 충돌)
- arXiv:2505.06120 — LLMs Get Lost In Multi-Turn Conversation (Laban et al.; 평균 성능 하락)
- arXiv:2505.15392 — Understanding the Anchoring Effect of LLM with Synthetic Data
- arXiv:2502.08177 — SycEval: Evaluating LLM Sycophancy
- arXiv:2504.00180 — Contradiction Detection in RAG Systems
- arXiv:2602.04288 — Contextual Drag: How Errors in the Context Affect LLM Reasoning
- arXiv:2603.03308 — Old Habits Die Hard: How Conversational History Geometrically Traps LLMs
- arXiv:2603.12123 — Cross-Context Review: Separating Production and Review Sessions
- arXiv:2606.20245 — Explicit Knowledge Conflict Resolution for LLM Inference
- arXiv:2511.04694 — Reasoning Up the Instruction Ladder for Controllable Language Models
- arXiv:2604.09443 — Many-Tier Instruction Hierarchy in LLM Agents
해결 · 메모리 · 멀티에이전트 (06 해결전략)
- arXiv:2310.08560 — MemGPT: Towards LLMs as Operating Systems
- arXiv:2510.00615 — ACON: Optimizing Context Compression for Long-horizon LLM Agents
- arXiv:2501.06322 — Multi-Agent Collaboration Mechanisms: A Survey of LLMs
- arXiv:2505.02279 — A Survey of Agent Interoperability Protocols (MCP, ACP, A2A, ANP)
거버넌스 · 멀티에이전트 보안 · provenance (07 거버넌스)
- arXiv:2212.08073 — Constitutional AI: Harmlessness from AI Feedback
- arXiv:2505.02077 — Open Challenges in Multi-Agent Security
- arXiv:2602.17913 — From Lossy to Verified: A Provenance-Aware Tiered Memory for Agents
- arXiv:2602.22724 — AgentSentry: Mitigating Indirect Prompt Injection via Temporal Causal Diagnostics
- arXiv:2603.11768 — Governing Evolving Memory in LLM Agents (SSGM Framework)
- arXiv:2603.18043 — The Provenance Paradox in Multi-Agent LLM Routing
- arXiv:2604.01664 — ContextBudget: Budget-Aware Context Management for Long-Horizon Search Agents
- arXiv:2604.07007 — AgentCity: Constitutional Governance for Autonomous Agent Economies via Separation of Power
- arXiv:2604.07911 — Dynamic Attentional Context Scoping (DACS)
- arXiv:2604.16339 — Semantic Consensus: Process-Aware Conflict Detection and Resolution
보강 추가 논문 (2026-06-20 — 커버리지·심화 확장)
압축·컨텍스트 확장·KV-캐시 (06_03-Compress전략-요약-프루닝-KV캐시최적화 · 05_03-메커니즘-왜-발생하는가)
- arXiv:2104.09864 — RoFormer: Enhanced Transformer with Rotary Position Embedding (RoPE)
- arXiv:2306.15595 — Extending Context Window of LLMs via Positional Interpolation
- arXiv:2309.00071 — YaRN: Efficient Context Window Extension of LLMs
- arXiv:2309.17453 — Efficient Streaming Language Models with Attention Sinks (StreamingLLM)
- arXiv:2304.08467 — Learning to Compress Prompts with Gist Tokens
- arXiv:2310.05736 — LLMLingua: Compressing Prompts for Accelerated Inference
- arXiv:2412.19442 — A Survey on LLM Acceleration based on KV Cache Management
- arXiv:2603.20397 — KV Cache Optimization Strategies for Scalable and Efficient LLM Inference
RAG vs 롱컨텍스트 (06_10-RAG-vs-롱컨텍스트-논쟁)
- arXiv:2407.16833 — Retrieval Augmented Generation or Long-Context LLMs? A Comprehensive Study and Hybrid Approach
에이전트 벤치마크 (06_11-에이전트-벤치마크-tau-swe-gaia)
- arXiv:2406.12045 — τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
- arXiv:2506.07982 — τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment
- arXiv:2310.06770 — SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
- arXiv:2311.12983 — GAIA: a benchmark for General AI Assistants
- arXiv:2307.13854 — WebArena: A Realistic Web Environment for Building Autonomous Agents
오염 실사고·다중에이전트 합의 (01_12-사례-산업-실사고 · 07_10-합의-알고리즘-비잔틴-정족수)
- arXiv:2509.10540 — EchoLeak: First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System (CVE-2025-32711)
- arXiv:2305.14325 — Improving Factuality and Reasoning in Language Models through Multiagent Debate
- arXiv:2502.19130 — Voting or Consensus? Decision-Making in Multi-Agent Debate
2. 공식 문서 (Anthropic · OpenAI) — 08 실전
Anthropic — Claude Code
- Memory (CLAUDE.md 계층) — https://code.claude.com/docs/en/memory
- Context window / management — https://code.claude.com/docs/en/context-window
- Costs / context — https://code.claude.com/docs/en/costs
- Sub-agents (컨텍스트 격리) — https://code.claude.com/docs/en/sub-agents
- Hooks — https://code.claude.com/docs/en/hooks
- Prompt caching — https://code.claude.com/docs/en/prompt-caching
- Anthropic, “Managing context” (context editing / memory tool, beta) — https://claude.com/blog/context-management
- Anthropic, “Building effective agents” (2024-12) — https://www.anthropic.com/engineering/building-effective-agents
- Anthropic, “How we built our multi-agent research system” — https://www.anthropic.com/engineering/multi-agent-research-system
OpenAI — Codex
- AGENTS.md — https://developers.openai.com/codex/guides/agents-md · (표준 사이트: https://agents.md)
- Memories — https://developers.openai.com/codex/memories
- Chronicle — https://developers.openai.com/codex/memories/chronicle
- Compaction — https://platform.openai.com/docs/guides/compaction
3. 산업 블로그 · 리포트
- Cognition, “Don’t Build Multi-Agents” (Walden Yan, 2025) — https://cognition.ai/blog/dont-build-multi-agents · 단일 선형 에이전트를 옹호하는 논거. 06 해결전략 · 07 거버넌스.
- Manus, “Context Engineering for AI Agents: Lessons from Building Manus” (Yichao ‘Peak’ Ji, 2025) — https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus · KV-캐시와 파일시스템 메모리.
- Letta, “Agent Memory” — https://www.letta.com/blog/agent-memory/ · “Sleep-time Compute” — https://www.letta.com/blog/sleep-time-compute
- Mem0, “State of AI Agent Memory 2026” — https://mem0.ai/blog/state-of-ai-agent-memory-2026
- EclipseSource, “MCP Context Overload” (2026-01) — https://eclipsesource.com/blogs/2026/01/22/mcp-context-overload/
- Demiliani, “MCP and the Too Many Tools Problem” (2025-09) — https://demiliani.com/2025/09/04/model-context-protocol-and-the-too-many-tools-problem/
- Leonie Monigatti, “Virtual context management with MemGPT and Letta” (2024) — https://www.leoniemonigatti.com/blog/memgpt.html
검증 메모
- arXiv 74편은
export.arxiv.org/api/query?id_list=…로 ID와 제목을 대조해 모두 실재함을 확인했다(2026-06-19). - 2026 보강분 22편도 같은 API로 대조해 확인했다(2026-06-20). 상세는 papers-2026-신규, 누적 96편.
- 2026 아티클 6편은 원문 URL을 직접 열어 확인했다. articles-2026.
- 2026 도구 신기능은 직접 확인한 것과 보도로만 접한 것을 나눠 tools-2026에 정리했다. MCP 2026-07-28 스펙은 미래 날짜라 제외했다.
- 2026년 논문(2601~2606) 일부는 아직 최신이라 결론이 확정적이지 않을 수 있다. 해당 노트 본문의 유보 표기를 함께 보라.
- 블로그와 공식 문서는 원문을 직접 확인해 인용했다. 다만 Databricks 등 2차 인용으로 들어온 구체 수치에는 “확인 권장” 표기를 남겨 두었다.