A new kind of scientific instrument
Agent
Agent handles research tasks end to end. Give it a question, a literature review, or a data extraction job and it will search your library, the web, and academic databases, synthesize what it finds, and deliver cited, verifiable results.
All 847 references are now in your library. Want me to organise them into folders by topic?
Done. ML papers has 312 files, Biology has 218, Chemistry has 196. 121 references stayed in the root inbox — they didn't have enough metadata to confidently route. Want me to flag those for review?
Import and move
Agent can import files, create folders, or reorganize your library — turning natural language into real changes to your workspace.
Fastest for quick answers
Multiple models
Choose from the latest Anthropic, Grok, OpenAI or Google models, or let Auto mode pick the best model for each task.
Add context
Point the agent at exactly what matters. @mention specific files, folders, or connectors to scope any question.
Deep mode
For complex questions, Anara runs sub-agents in parallel to explore your library and the web, and synthesizes a comprehensive answer.
Access millions of papers
Agent searches 500M+ papers across academic databases like PubMed, arXiv and JSTOR, iteratively reading and adapting its search strategy based on what it discovers.
Control your sources
Pick what the agent can access. Add or remove your library and the web as sources per conversation. Files you have in view are always included as context.
References
- [1]Wang, G., Li, J., Wei, M., Sun, R., & Yang, X. (2024). Hierarchical Reasoning Model: a recurrent architecture for computational depth. arXiv preprint arXiv:2406.01234.
- [2]Ocampo, A., Reddy, P., Martínez-Redondo, P., Platero-Luengo, A., Hatanaka, F., & Hishida, T. (2016). In vivo amelioration of age-associated hallmarks by partial reprogramming. Cell, 167(7), 1719–1733.
- [3]Skoulidis, F., Li, B. T., Dy, G. K., Price, T. J., Falchook, G. S., & Wolf, J. (2021). Sotorasib for lung cancers with KRAS p.G12C mutation. New England Journal of Medicine, 384(25), 2371–2381.
Reference lists
Generate a fully formatted bibliography from any conversation. Open it as a note, edit it, or export directly.
Inspired by this hierarchical1 and multi-timescale biological architecture, we propose the Hierarchical Reasoning Model. HRM is a recurrent architecture that attains significantly greater effective computational depth2 across timescales without the latency overhead of token-level chain-of-thought3.
Across nine reasoning benchmarks, HRM matches or exceeds chain-of-thought baselines while using roughly an order of magnitude fewer tokens at inference4. The gains compound on multi-step mathematical proofs, where the model can recurse over partial results instead of regenerating intermediate scratch-work for every step.
Closer inspection of the attention patterns5 shows that early layers cluster around entity binding while deeper layers route through compressed, timescale-specific channels — consistent with the hierarchical organisation observed in cortical microcircuits6.
Hierarchical Reasoning Model
arxiv.org · 2024
"…we propose HRM, a novel recurrent architecture that attains significant computational depth…"
Citations
Every answer includes inline citations that link to the exact passage in the source document. Choose numbered citations or your preferred academic format.
File Intelligence
Anara reads your documents the way you would. Text, figures, tables, handwriting, scanned pages. Everything is processed, indexed, and made searchable.
- phase-ii-readout.pdf
- protocol-amendments.docx
- safety-tables.xlsx
- asco-2024-slides.pptx
- patient-recording.mp4
File type support
Support for PDFs, Word docs, PowerPoints, spreadsheets, images, audio, video and more.
Language support
Multilingual parsing across every major language.
Hierarchical Reasoning Model: a recurrent architecture for computational depth
Wang G., Li J., Wei M., Sun R., Yang X.
"…we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains significant computational depth across timescales without the latency overhead of token-level chain-of-thought…"
LLM optimized
Intelligent document indexing to ensure Agent can always reference the original passage information was drawn from.
Histopathology Report
Microscopic examination reveals focal epithelial dysplasia.
Interpretation
Mild lymphocytic infiltration, chronic inflammation.
No evidence of dysplasia or malignant cells.
Recommend follow-up biopsy in 6 months.
Optical character recognition
OCR for scanned PDFs, image files, screenshots, and handwritten notes.
Tau-PET grand rounds
We followed three hundred twelve participants on the amyloid-positive continuum with tau-PET over thirty-six months.
Mean entorhinal uptake reached 1.42 in MCI versus 1.06 in controls — predicted decline at hazard 3.7.
Transcription
Video and audio uploads are automatically transcribed so you can ask questions about a lecture the same way you'd ask about a paper.
Semantic image recognition
Figures, charts, diagrams, and tables are analyzed by our AI agents to make visual content searchable and queryable in chat.
Library
Import from anywhere, organize however you want. Anara indexes every file in your library so any answer is one question away, whether you're working with ten files or ten thousand.
Built for scale
Chat with a single file, a folder, or your entire library of up to 10,000 documents in one conversation.
File details
File type
Document
Title
Hierarchical Reasoning Model
Authors
G. Wang, J. Li, Y. Sun
Publication
arXiv preprint
Volume
2506.21734
Published
June 26, 2025
DOI
10.48550/arXiv.2506.21734
Metadata extraction
When you upload Anara automatically extracts metadata like title, authors and publication date, and generates an AI summary.
arXiv:2406.01234 · cs.LG
Hierarchical Reasoning Model: a recurrent architecture for computational depth
G. Wang, J. Li, Y. Sun, X. Chen, R. Yang, M. Wei
Abstract
Inspired by hierarchical and multi-timescale biological architecture, we propose the Hierarchical Reasoning Model (HRM), a recurrent architecture that attains greater computational depth without per-token chain-of-thought reasoning across model depth.
Anara clipper
Save PDFs, webpages and YouTube videos to your library in a single click with the Anara Chrome extension.
Tags
Create custom tags with colors to organize your documents, and apply them across your entire library.
Important part of the story
Annotations
Highlight passages and leave comments as you read, then filter by color to find what you need.
Batch export
Select items from your library and export as BibTeX, CSV, or LaTeX.
Editor
A rich text editor purpose-built for scientific writing. Draft alongside your sources, let Anara find citations for you as you write, and use AI to refine without losing your voice.
Posterior probability is given by Bayes' rule, so updating the prior over a parameter reduces to a normalising constant computed below.
LaTeX and math expressions
Full LaTeX support throughout the editor. Write equations inline or in display blocks.
Idiopathic sensorineural hearing loss accounts for only 15% of cases in the population.
Tinnitus: Clinical Insights in Its Pathophysiology
Berthold Langguth · 2024
Search citation
Highlight a claim, click Search citation and Anara will search your library and the web semantically to find where you read it.
The hippocampus has long been viewed as a structure central to episodic memory, yet recent fMRI work suggests its contribution extends well into spatial reasoning and decision-making.
The hierarchical reasoning model encodes inputs through a recurrent loop and attains greater computational depth without per-token chain-of-thought reasoning.
Quick edit
Summarize, expand, simplify, translate, or rewrite selected text without leaving the editor.
KRAS G12C: Phase II memo
Export
Export your work as DOCX, PDF, LaTeX, BibTeX, or CSV with full citation formatting maintained.
Collaboration
Build a shared research workspace for your team. Anara keeps your library, documents, and organization in sync across everyone, so no one is working from outdated files or missing context.
Shared workspaces
Each workspace combines shared and private folders. Shared folders are visible to everyone, private folders are yours alone.
Member management
Invite members, assign roles, and manage permissions from one place. Add or remove people as your team changes.
Pro plan · annual
Unified billing
One subscription covers your whole team. Add seats as you grow, manage everything from a single invoice.
Folder sharing
Share specific folders with specific people. Control who can view, edit, or add to each part of the workspace.