General Agentic Memory uses dual-agent design, beats RAG on benchmarks
Why does memory matter for today’s conversational AI? As models churn through longer dialogues, they often lose track of earlier turns—a problem researchers label “context rot.” The new General Agentic Memory system claims to curb that decay and even surpass Retrieval‑Augmented Generation (RAG) on established memory benchmarks. While many approaches simply pull in external documents at query time, this work pushes the idea of persistent, structured recall.
The authors report that their method keeps a running record of interactions, then surfaces relevant facts without re‑processing the entire transcript each time. That sounds promising, but the real test lies in how the architecture splits responsibilities and whether the split yields measurable gains. Below, the paper describes the core of that design—a dual‑agent setup that separates summarization from deeper research.
**A dual-agent architecture**
A dual-agent architecture GAM uses a dual architecture consisting of two specialized components: a "Memorizer" and a "Researcher." The Memorizer runs in the background during interactions. While it creates simple summaries, it also archives the full conversation history in a database called the "page store." It segments the conversation into pages and tags them with context to make retrieval easier. The Researcher activates only when the agent receives a specific request.
Instead of simply looking up memory, it conducts "deep research"--analyzing the query, planning a search strategy, and using tools to dig through the page store. It uses three methods: vector search for thematic similarities, BM25 search for exact keywords, or direct access via page IDs. The agent verifies its search results and reflects on whether the information is sufficient.
If necessary, it starts new queries before generating an answer. Outperforming RAG and long-context models The team tested GAM against conventional methods like Retrieval-Augmented Generation (RAG) and models with massive context windows like GPT-4o-mini and Qwen2.5-14B. According to the paper, GAM beat the competition in every benchmark.
The gap was widest in tasks requiring information linking over long periods. In the RULER benchmark, which tracks variables over many steps, GAM hit over 90 percent accuracy while conventional RAG approaches and other storage systems largely failed. The researchers believe GAM succeeds because its iterative search finds hidden details that compressed summaries miss.
The system also scales well with compute: allowing the Researcher more steps and reflection time further improves answer quality.
The paper shows that General Agentic Memory can curb context rot, a persistent flaw in today’s agents. By pairing a background “Memorizer” that drafts brief summaries with a “Researcher” that pulls from a full‑conversation database—named “pag”—the system keeps detail alive beyond the usual window limits. Benchmarks indicate GAM outperforms Retrieval‑Augmented Generation on the memory tests the authors selected.
Yet the experiments focus on controlled settings; it’s unclear whether the dual‑agent pipeline will hold up under open‑ended, real‑world dialogues. The researchers’ Chinese‑Hong Kong team emphasizes compression and archival as core to the design, but they do not disclose how the architecture scales with larger models or longer histories. In short, the approach offers a concrete step toward more persistent AI memory, though further validation will be needed to confirm its broader applicability.
Further Reading
- General Agentic Memory Via Deep Research - arXiv
- General Agentic Memory (GAM) Overview - Emergent Mind
- General Agentic Memory tackles context rot and outperforms RAG in memory benchmarks - The Decoder
- General Agentic Memory Via Deep Research - alphaXiv - alphaXiv
- General Agentic Memory (GAM) - Jimmy Song
More in Research & Benchmarks
ARC benchmark declines as labs tune AI to optimize its specific logic
The ARC benchmark, once a yardstick for measuring an AI’s ability to reason abstractly like a...
NeurIPS 2025: Top 4 Papers Highlight Shift From Bigger Models to Limits
The four papers that stole the spotlight at NeurIPS 2025 aren’t shouting about ever‑larger...
97% Can't Distinguish AI Music; 71% Surprised, 51% Uncomfortable
A recent benchmark put AI‑generated tracks head‑to‑head with human‑made songs and found that 97...
TPUs Designed for Deep Learning Can Outperform GPUs in Many Workloads
When I first looked at the hardware people pick for today’s AI jobs, the first thing that struck me...
Wipro partners with IISc and FSID for AI and quantum research collaboration
Right now Indian companies seem to be reaching out more to university labs for projects that could...
Common Questions Answered
What specific issue in conversational AI does General Agentic Memory aim to address?
General Agentic Memory targets the problem of "context rot," where models lose track of earlier dialogue turns as conversations grow longer. By maintaining a persistent, structured recall of past interactions, the system seeks to keep relevant context alive beyond the model's native window limits.
How does the dual‑agent architecture of GAM work, and what are the responsibilities of the Memorizer and the Researcher?
The GAM architecture splits functionality between two specialized agents: the Memorizer runs continuously in the background, creating brief summaries and archiving the full conversation into a database called the page store. The Researcher is invoked only on explicit requests, pulling detailed information from the page store to answer queries that require deeper context.
What is the purpose of the "page store" in GAM, and how does it improve retrieval compared to standard Retrieval‑Augmented Generation?
The page store organizes the conversation into segmented "pages" and tags each segment with contextual metadata, making it easier to locate relevant information later. This structured indexing allows the Researcher to retrieve precise excerpts quickly, whereas traditional Retrieval‑Augmented Generation typically relies on ad‑hoc document retrieval at query time.
According to the paper, how does General Agentic Memory perform on memory benchmarks relative to Retrieval‑Augmented Generation?
Benchmarks reported in the study show that GAM outperforms Retrieval‑Augmented Generation on the selected memory tests, demonstrating a lower rate of context loss and higher accuracy in recalling past dialogue details. However, the authors note that these experiments were conducted in controlled settings, leaving open questions about real‑world scalability.