On November 9, 2026, researchers from the City University of Hong Kong and Tencent Jarvis Lab released a preprint titled “Can LLM Agents Organize Their Own Memory?” (arXiv:2611.11929), introducing SoMem, an open-source framework that enables autonomous large language model agents to structure persistent memory as editable, linked Markdown documents rather than static chunked vector databases.
Autonomous Memory Organization Through Linked Documents
As large language models are increasingly deployed in long-horizon scientific workflows, project coordination, and autonomous research assistance, their inability to update weights during deployment creates a critical dependency on external memory. Traditional architectures generally follow two paradigms: rigid relational databases with fixed schemas, or unstructured chunk-based vector stores queried via semantic retrieval. Both exhibit severe operational bottlenecks in extended multi-session tasks. Fixed schemas constrain how concepts can evolve as new observations arrive, while flat vector stores scatter interdependent multi-hop facts across disconnected chunks without preserving causal or chronological structure.
In the preprint published on arXiv by lead authors Zichuan Fu, Junjia Qi, and colleagues from City University of Hong Kong alongside Tencent Jarvis Lab, the research team investigates whether an autonomous agent can manage its own memory lifecycle using ordinary documents. Under the proposed SoMem (Self-Organized Memory) paradigm, the agent actively decides what information to summarize, how to cluster related observations into topical files, when to split unwieldy documents, and how to restructure existing topics as new evidence accumulates.
A central feature of SoMem is its strict evidence preservation contract. Rather than replacing raw interaction logs with lossy summaries, the system archives every dialogue turn into an append-only, immutable Source store. When the agent generates or updates semantic summaries in Topic files, every statement is explicitly tagged with block identifiers and footnote-style links pointing directly to the original Source turns. This design prevents memory drift and hallucinated synthesis, allowing downstream tasks to verify factual claims against ground-truth evidence.
Architectural Core: Validated Updates and Derived Multi-View Indexes
The SoMem architecture divides the memory lifecycle into distinct writing, indexing, and reading phases governed by runtime guarantees:
- Immutable Source Records: An append-only repository that archives raw user interactions and tool execution logs verbatim. The agent can read and cite Source records, but runtime controls strictly forbid modification or deletion.
- Editable Topic Documents: A dynamic directory of Markdown files where the agent maintains thematic knowledge. Documents are structured using headings, with individual paragraphs serving as addressable memory units tagged with persistent block identifiers. Inter-document references allow paragraphs to cross-reference related concepts across the workspace.
- Size-Bounded Core Memory: A dedicated, bounded Markdown file that holds high-priority persistent context, including user preferences and operational requirements, automatically injected into every query’s initial context window.
- Deterministic View Derivation: To avoid having the LLM maintain separate chronological, relational, and recency indexes, the runtime automatically derives three complementary access views directly from the Topic files:
- Relational Graph View: Parses block-level cross-references into directed link and backlink indices, enabling bi-directional traversal across conceptual relationships without dedicated knowledge graph extraction pipelines.
- Temporal Timeline View: Automatically extracts timestamp metadata from evidence annotations, assembling chronological documents categorized by year, month, and day (`timeline/YYYY/MM/DD.md`) to support temporal reasoning.
- Recent Buffer: Tracks unit creation timestamps to maintain a rolling window of the most recent memory operations, ensuring immediate conversational continuity.
- Staged Runtime Validation: Proposed changes are staged in an isolated workspace. The runtime verifies that Source logs remain untouched, paragraph syntax complies with formatting contracts, internal links resolve to valid block targets, and workspace path boundaries are respected before executing an atomic install.
The following table summarizes how SoMem compares against conventional agent memory architectures across structural dimensions:
| Architectural Dimension | Flat Chunked Vector Stores | Predefined Schema Databases | SoMem Linked Documents |
|---|---|---|---|
| Storage Format | Isolated text chunks and vector embeddings | Relational tables or key-value JSON fields | Editable Markdown documents with block IDs |
| Topic Reorganization | None (chunks remain static once embedded) | Rigid (schema modifications require re-indexing) | Dynamic (agent creates, splits, and merges topics) |
| Evidence Traceability | Approximate (cosine similarity scores) | Manual foreign key linkages | Footnote pointers to immutable Source logs |
| Index Derivation | Approximate nearest-neighbor graphs | Database secondary indices | Deterministic Graph, Timeline, and Recent views |
| Update Integrity | Unchecked chunk insertions | Schema type constraints | Staged runtime validation and atomic commits |
Empirical Benchmark Evaluation and Cross-Model Portability
The research team evaluated SoMem across three long-term conversational memory benchmarks: LoCoMo (Maharana et al., 2024), BEAM-100K (Tavakoli et al., 2025), and LongMemEval-S (Wu et al., 2025). The primary benchmark suite tested all memory systems using a standardized setup with GPT-6 Luna serving as both the memory-construction Writer and question-answering Reader, evaluated by a GPT-4o-mini judge.
Across all three benchmarks, SoMem established new state-of-the-art results, surpassing leading specialized memory baselines including E-mem, Mem0, LightMem, Nemori, MemoryOS, and A-Mem, as well as direct full-context prompting:
- LoCoMo (1,540 category 1-4 questions): SoMem attained 88.96% overall accuracy, outperforming the strongest baseline E-mem (86.30%) by 2.66 percentage points, Mem0 (80.32%) by 8.64 percentage points, and Full Context (84.22%) by 4.74 percentage points. SoMem showed notable strength in multi-hop reasoning (83.33%) and temporal questions (88.16%).
- BEAM-100K (400 questions over 20 conversations): SoMem achieved a mean nugget score of 54.02% and a pass rate of 61.50%, beating Full Context (50.92% score, 55.25% pass) by 3.10 percentage points and A-Mem (50.51% score, 55.75% pass) by 3.51 percentage points.
- LongMemEval-S (500 questions): SoMem scored 76.50% overall accuracy, surpassing LightMem (74.50%) by 2.00 percentage points, Mem0 (72.80%) by 3.70 percentage points, and Full Context (65.80%) by 10.70 percentage points.
The paper also examined cross-model portability, demonstrating that a memory workspace constructed by one model can be frozen and queried by different Reader models. In cross-model configurations reported in the technical paper on the SoMem GitHub repository and paper preprint on Hugging Face Papers, pairing a GPT-5.6 Luna Writer with a Qwen 3.8 Flash Reader achieved 93.25% accuracy on LoCoMo, while an end-to-end DeepSeek V4 Flash setup reached 90.20% on LongMemEval-S. Furthermore, empirical scaling experiments indicated that dedicating additional reasoning effort to the Reader model not only improves answer accuracy on complex multi-hop queries, but can also reduce net token consumption through more targeted document reads.
Broader Landscape and Technical Reception
The release of SoMem reflects a broader architectural shift in autonomous agent engineering away from opaque vector embeddings and toward transparent, inspectable text workspaces. While early agent memory systems relied heavily on dense vector databases or isolated scratchpads, recent developments emphasize structural legibility and human-in-the-loop inspectability.
Independent research roundups, including coverage on the SoMem project site and technical commentary across community platforms, highlight that storing memory in standardized Markdown formats allows human operators, auditors, and external verification tools to directly inspect an agent’s internal state. Because memories exist as readable text files rather than high-dimensional vectors, errors can be audited or corrected without retraining models or re-indexing entire vector databases.
What It Means for Researchers, Universities, and Research Offices
The emergence of self-organizing, evidence-linked agent memory carries concrete practical implications for university research administrators, principal investigators, and computational laboratories:
- Auditability and Verification: Autonomous agents assisting in literature synthesis, data curation, or grant drafting must produce verifiable claims. By anchoring every summary to immutable Source turns, SoMem directly supports rigorous AI output verification, allowing research integrity officers to trace claims back to primary empirical records.
- Reproducible Scientific Provenance: Maintaining memory workspaces as standard text files allows researchers to track memory evolution using standard version control systems like Git. This structure provides tamper-evident provenance trails compatible with emerging institutional frameworks like C2PA content provenance and machine-actionable data governance.
- Long-Horizon Research Capability: Long-term academic investigations require agents to maintain operational consistency over weeks or months of experimentation. The benchmark findings in SoMem connect directly to capability evaluations such as the METR time horizon framework, demonstrating that structured memory organization significantly mitigates context degradation in extended tasks.
- Transparent Disclosure Standards: As publishers and funding agencies establish stricter disclosure requirements for generative technologies, maintaining explicit operational traces facilitates compliance with AI transparency marker reporting protocols.
What Happens Next
The authors have made the complete SoMem codebase, evaluation harnesses, and documentation publicly accessible under an open-source license on GitHub. Looking forward, the research team identified several planned enhancements, including scaling evaluations to the million-token BEAM-1M benchmark suite, expanding tool integrations for interactive command-line environments, and formalizing cryptographic verification protocols for immutable source archives. Continued validation by third-party research teams will help establish whether self-organizing document trees can serve as a universal substrate for multi-agent scientific collaboration.
Sources
- arXiv: Can LLM Agents Organize Their Own Memory? (Preprint Abstract) — arXiv (Cornell University), 2026-11-09.
- arXiv: Can LLM Agents Organize Their Own Memory? (Full Technical Report PDF) — arXiv (Cornell University), 2026-11-09.
- GitHub: SoMem Open-Source Repository and Documentation — City University of Hong Kong and Tencent Jarvis Lab, 2026-11-10.
- Hugging Face Papers: Can LLM Agents Organize Their Own Memory? — Hugging Face, 2026-11-10.
- SoMem Project Page and Overview — City University of Hong Kong, 2026-11-10.







