Researchers at City University of Hong Kong and Tencent Jarvis Lab report that AI agents scored higher on three long-term memory tests when they organized their own memory as linked documents. In a paper posted to arXiv on 9 November, their method, SoMem, topped six named memory systems, Mem0 among them, and also topped feeding the model the full conversation history.
On the LoCoMo benchmark, SoMem scored 88.96% to Mem0’s 80.32%, a gap of 8.64 points by our arithmetic. The paper’s margins over the strongest baseline on each test are 2.66, 3.10 and 2.00 percentage points. All scores come from the authors’ own runs, graded by an AI judge.
Business Pill · RAG Explained
A short explainer on how a model answers from your documents. It teaches the general idea only and does not discuss the SoMem paper.
The key insight: As we read the paper, the gain came less from clever organization than from keeping the evidence: SoMem scored 88.96% on LoCoMo, and without its links back to the original conversations it scored 80.32%, the same as Mem0, in the authors’ own runs.
Existing approaches, the abstract says, “often rely on predefined records and fixed organizational rules”. SoMem hands that job to the agent. In the abstract’s words, “The agent decides what to summarize, how to group related information, and when to revise existing topics, and every summary it produces stays linked to the original interactions as evidence”.
The software around the model does the bookkeeping: “The runtime validates updates and automatically derives temporal, relational, and recent-memory views from the organized content”. When a question arrives, “the agent chooses what to search and read, following document links and inspecting original evidence as needed”. The memory is stored as ordinary Markdown documents, and the code is public on GitHub.
Overall accuracy on LoCoMo’s 1,540 questions, with GPT-6 Luna writing the memory and answering for every method, per Table 2 of the SoMem paper. Graded by an AI judge in the authors’ own runs; the 8.64-point gap to Mem0 is our arithmetic.
What Was Posted on X
Super interesting work on agent memory.
(this is a must-read)
It discusses whether agents can manage their own memory. If so, what would that self-organized memory look like?
This is the closest implementation that I have seen to my own memory system. I mostly use linked documents with my agents, and it has been working extremely well. It works well because frontier models have gotten better at persisting external knowledge.
The problem:
Memory systems usually fix the schema in advance, with set record types and rules for how to group them.
The solution:
SoMem gives the agent ordinary linked documents instead. The agent decides what to summarize, which topics to create, and when to revise them.
Every summary links back to the original interaction as evidence, and the runtime builds time, relation, and recency views on top.
At question time, the agent chooses what to search and follows links to the source when it needs to check.
SoMem beat the strongest baseline by 2.66 points on LoCoMo, 3.10 on BEAM-100K, and 2.00 on LongMemEval-S. Memory built by one model also worked when read by another.
If your agent already reads and writes files, try letting it organize memory as linked notes before adding a vector store.
Elvis Saravia (@omarsar0) posted this at 14:37 UTC on 12 November 2026, summarising the SoMem paper and linking to it on arXiv. Gennaro passed it to the desk. The piece quotes its closing advice; every score comes from the paper’s own tables.
The Results, Test by Test
The paper runs every method with the same model, GPT-6 Luna, writing the memory and answering the questions. On LoCoMo’s 1,540 questions SoMem scored 88.96%, against 86.30% for E-mem, the next best, and 84.22% for the full history with no memory system.
On LongMemEval-S’s 500 questions SoMem scored 76.50%, against 74.50% for LightMem and 72.80% for Mem0; the full history scored 65.80%, 10.70 points lower by our arithmetic. On BEAM-100K’s 400 questions, SoMem’s mean score was 54.02 against 50.92 for the full history, the strongest baseline there.
Memory built by one model worked when another read it. Memory written by GPT-5.6 Luna scored 93.25% on LoCoMo when Qwen 3.8 Flash answered from it, against 92.34% when GPT-5.6 Luna read its own notes.
Why the Links Matter
The authors switched parts off one at a time. Removing the links back to the original conversations, which the paper calls Source, cut LoCoMo accuracy from 88.96% to 80.32%. By our reading of the tables, that is the same score Mem0 posted. Removing the topic documents cut it to 86.04%.
The other parts, a timeline, a link graph, a recent-memory preload and a small core file, each moved overall accuracy by at most 1.30 points, within sampling uncertainty, the paper says.
LoCoMo overall accuracy with one SoMem component removed at a time, per Table 5 of the paper (single runs, authors’ own). 80.32 is also Mem0’s score in Table 2; the 8.64-point drop is our arithmetic.
The Business Angle: the Memory Layer
Memory has become its own layer in agent software. The paper benchmarks six memory systems, Mem0 among them, and cites LangChain’s LangMem tools. LangChain’s chief executive, Harrison Chase, told the 20VC podcast in October that “One thing about agents that is rare today but will be incredibly common in the next few years is this memory layer”.
Elvis Saravia, who flagged the paper on X, wrote: “If your agent already reads and writes files, try letting it organize memory as linked notes before adding a vector store.”
The Structural Read
As we read the cross-model table, memory written by one model could be read by another without rebuilding it: all 12 Writer-Reader pairings in Table 6 scored between 90.13% and 93.25% on LoCoMo, which makes the stored documents a reusable asset in this design.
As we read the method, SoMem needs files and links rather than a dedicated memory database, which puts it close to how agent harnesses already work with code and documents.
Fu et al., Can LLM Agents Organize Their Own Memory?, arXiv 2611.11929, abstract
“These findings demonstrate the feasibility of agents organizing and using their own persistent memory through ordinary documents.”
Three Implications
FOR AGENT BUILDERS The paper reports that agent-organized linked documents scored higher than six memory systems and the full history on three benchmarks, in the authors’ runs with an AI judge.
FOR MEMORY PRODUCTS The ablation the paper reports puts most of the gain on keeping links to the original conversations, a design choice any memory layer can be compared on.
FOR OUR COVERAGE We will add independent reproductions or production results if the authors or others publish them.
The framework puts it this way: “Orchestration, memory, tool use, and workflow execution have become the new sources of competitive advantage.”
As we read it, SoMem is a research result on the memory part of that list. The framework predates the paper and does not discuss it.
What Is Not Established
The paper’s own limits section says the tests cover question answering over conversation histories, not long-term personalization or task execution. “Evaluation relies on LLM judges without systematic human calibration and may be affected by judge biases”, it says. The paper’s token measurements exclude the cost of building the memory.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
An eight-author team from City University of Hong Kong and Tencent Jarvis Lab let agents organize their own memory as linked Markdown documents and scored higher than six memory systems and the full history on three tests, by the authors’ own grading. The links back to the original conversations did most of the work: without them, the score fell to Mem0’s.
A note on sourcing. We read the SoMem paper in full (arXiv 2611.11929, first version posted 9 November 2026) and its abstract page. All benchmark scores are the authors’ own runs, graded by an AI judge; we have not run the code. Elvis Saravia’s X post of 12 November flagged the paper. The LangChain and Anthropic context comes from our earlier pieces. Point gaps other than the paper’s 2.66, 3.10 and 2.00 are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
