agent memory long-context retrieval

Paper summary - General Agentic Memory Via Deep Research

Posted on December 01, 2025 at 12:50 PM

Paper summary: General Agentic Memory Via Deep Research

📚 Research topic and objective

  • The paper addresses the problem of memory management for AI agents (e.g. agents based on large language models, LLMs) — specifically how to store, recall, and reuse past work or context when agents run long multi-step tasks (e.g. multi-step reasoning, research, tool use). (arXiv)
  • Classical memory systems often build a static, compressed memory ahead-of-time (AOT). But that leads to loss of detail and poor adaptability if the future requests require fine-grained or unexpected info. (arXiv)
  • The objective: propose a new memory framework — General Agentic Memory (GAM) — that retains full history in a retrievable store, while summarizing key information in light memory; and when a request comes, dynamically “deep-researches” the store to build an optimized context tailored to that request. (arXiv)

🧠 What is GAM — how it works

GAM uses a two-part structure:

  • Memorizer (offline / as history accumulates):

    • When a new “session” (i.e. a unit of the agent’s past activity) arrives, the Memorizer produces a concise “memo” summarizing the important points. (arXiv)
    • At the same time, it stores the full session (with context header) into a “page-store” (like a database / library of full history), so no raw information is lost. (arXiv)

  • Researcher (online, when a request arrives):

    • Given the client’s request + the lightweight memory, the Researcher plans what information is needed, then searches the page-store (using retrieval tools such as vector search, keyword search, or ID-based lookup). (arXiv)
    • It integrates retrieved pages, reflects whether the result already suffices; if not, iterates — possibly retrieving more pages — until enough relevant context is collected. (arXiv)
    • Returns to the client a custom context: small (efficient) but containing the right detailed info to perform the task. (arXiv)

In effect: GAM trades some offline/online computation and retrieval effort for lossless memory + flexible, request-specific context rather than lossy, static summaries. (arXiv)

They further formalize the design as an optimization problem: produce a context that is “as small as possible” while maximizing downstream task performance. (arXiv)

✅ Key findings and conclusions

From their experiments (on multiple benchmarks), the authors find:

  • GAM consistently outperforms both memory-free methods (e.g. using a long-context LLM directly, or simple retrieval-augmented generation (RAG)) and existing memory-based methods (which rely on static memory) on a broad range of tasks. (arXiv)
  • The improvement is especially large on multi-hop reasoning tasks requiring tracking and integrating information dispersed across long contexts (e.g. in benchmarks such as HotpotQA and RULER). For example, on RULER’s multi-hop tracing (MT) tasks, GAM achieves over 90% accuracy, while many baselines fail. (arXiv)
  • GAM’s performance remains stable across different input-context lengths, showing that it scales well even when history/context becomes very large. (arXiv)
  • The authors show that just increasing LLM context window (e.g. using “long-LLM”) is not enough — large context can still lead to degraded performance due to distraction by irrelevant content (a phenomenon known as “context rot”). (arXiv)
  • Ablation studies: combining multiple retrieval tools (embedding search + keyword + page-id) yields better performance than any single tool, and both the memorizer and researcher modules are essential. Using only memory (without researcher) or only researcher (without memory) dramatically degrades performance. (arXiv)
  • Finally, efficiency: GAM’s time cost (offline + online) remains comparable with other memory-based systems, and significantly better cost-effectiveness overall: i.e. the benefit in task performance justifies the extra computation. (arXiv)

Conclusion: GAM appears to provide an effective, general-purpose memory architecture for LLM-based agents — one that preserves full historical information, yet delivers concise, optimized context on demand; this yields better task performance (especially for long-context and multi-step reasoning) and scales well in both length and complexity. (arXiv)

🔢 Critical data & facts (especially experimental results)

  • On the conversational memory benchmark (LoCoMo), across tasks (single-hop, multi-hop, temporal reasoning, open-domain), GAM achieves F1/BLEU-1 scores like 57.75 / 52.10 (etc.), noticeably higher than baselines such as long-LLM, RAG, MemoryOS, LightMem, etc. (arXiv)

  • On long-context reasoning benchmarks:

    • On HotpotQA (with large context: 56 K, 224 K, 448 K tokens), GAM obtains F1 scores like: 63.22, 64.56, 59.81 respectively. (arXiv)
    • On RULER (128 K tokens), in the multi-hop tracing (MT) task: GAM gets ≈ 93.2% (or “over 90%”) accuracy. (arXiv)
    • On narrative-level tasks (e.g. NarrativeQA), GAM also significantly outperforms baselines, showing robustness over very long inputs (average ≈ 87 K tokens in their test subset). (arXiv)
  • Regarding module sensitivity: using a smaller LLM for the researcher degrades performance significantly; the memory module is less sensitive to size reduction. (arXiv)
  • Regarding efficiency: For HotpotQA tasks, total processing time (offline build + online serve) for GAM is comparable to other memory-based systems (e.g. similar order to “Mem0” or “MemoryOS”), and much faster than some heavy offline-compression methods. (arXiv)

🧑‍🏫 What this means (for non-experts)

  • For AI agents that need to handle long histories, complex multi-step reasoning, or long documents — using a memory system like GAM helps them remember important details without losing information, even if the history is huge.
  • Instead of pre-compressing everything into a static summary (which might miss details needed later), the system keeps a full “library” + a lightweight index, and only pulls up what is relevant when needed. This is more flexible and powerful, especially when the agent faces unexpected or complex requests.
  • The approach appears general and domain-agnostic, meaning such a memory system could benefit many kinds of AI-agent applications: research assistants, code generation, decision support, long-run dialogues, etc.
  • Also suggests that simply giving LLMs a huge context window isn’t enough: you need smart memory management — storing raw material plus ability to dynamically retrieve relevant context.
FEATURED TAGS
computer program javascript nvm node.js Pipenv Python 美食 AI artifical intelligence Machine learning data science digital optimiser user profile Cooking cycling green railway feature spot 景点 e-commerce work technology F1 中秋节 dog setting sun sql photograph Alexandra canal flowers bee greenway corridors programming C++ passion fruit sentosa Marina bay sands pigeon squirrel Pandan reservoir rain otter Christmas orchard road PostgreSQL fintech sunset thean hou temple in sungai lembing 海上日出 SQL optimization pieces of memory 回忆 garden festival ta-lib backtrader chatGPT generative AI stable diffusion webui draw.io streamlit LLM speech recognition AI goverance prompt engineering fastapi stock trading artificial-intelligence Tariffs AI coding AI agent FastAPI 人工智能 Tesla AI5 AI6 FSD AI Safety AI governance LLM risk management Vertical AI Insight by LLM LLM evaluation AI safety enterprise AI security AI Governance Privacy & Data Protection Compliance Microsoft Scale AI Claude Anthropic 新加坡传统早餐 咖啡 Coffee Singapore traditional coffee breakfast Quantitative Assessment Oracle OpenAI Market Analysis Dot-Com Era AI Era Rise and fall of U.S. High-Tech Companies Technology innovation Sun Microsystems Bell Lab Agentic AI McKinsey report Dot.com era AI era Speech recognition Natural language processing ChatGPT Meta Privacy Google PayPal Edge AI Enterprise AI Nvdia AI cluster COE Singapore Shadow AI AI Goverance & risk Tiny Hopping Robot Robot Materials SCIGEN RL environments Reinforcement learning Continuous learning Google play store AI strategy Model Minimalism Fine-tuning smaller models LLM inference Closed models Open models Privacy trade-off MIT Innovations Federal Reserve Rate Cut Mortgage Interest Rates Credit Card Debt Management Nvidia SOC automation Investor Sentiment Enterprise AI adoption AI Innovation AI Agents AI Infrastructure Humanoid robots AI benchmarks AI productivity Generative AI Workslop Federal Reserve Enterprise AI Adoption AI automation Multimodal AI Google AI AI agents AI integration Market Volatility Government Shutdown Rate-cut odds AI Fine-Tuning LLMOps Frontier Models Hugging Face Multimodal Models Energy Efficiency AI coding assistants AI infrastructure Semiconductors Gold & index inclusion Multimodal Chinese open-source AI AI hardware Semiconductor supply chain Open-Source AI prompt injection LLM security red teaming AI spending AI Bubble Quantum Computing Open-source AI AI shopping Multi-agent systems AI research breakthroughs AI in finance Financial regulation Custom AI Chips Solo Founder Success Newsletter Business Models Indie Entrepreneur Growth Apple Claude AI Infrastructure AI chips robotaxi Gemini AI Global expansion AI security embodied AI AI tools IPO artificial intelligence venture capital multimodal AI startup funding AI chatbot AI browser space funding Alibaba quantum computing DeepSeek enterprise AI AI investing tech bubble reinforcement learning AI investment prompt injection attacks AI red teaming agentic browsing agentic AI cybersecurity AI search AI boom AI adoption data centre multimodal models model quantization AI therapy neuro-symbolic AI AI bubble tech valuations sovereign cloud Microsoft Sentinel large language models vision-language model open-source LLM Digital Assets Qwen3‑Max AI drug discovery open-source AI Hugging Face updates Gemini 3 investment-grade bonds data residency AI funding AI regulation Gemini 3 AI banking GPT-5.2