Logically vs. Agent Bayes: Which Fits Your Research Workflow?

First, which Logically?

Two products share the name. This page is about logically.app, the AI research and writing workspace that used to be called Afforai. It is not Logically AI at logically.ai, the UK company working on misinformation and threat intelligence, and it is not an argument-mapping tool. If you arrived looking for either of those, this is the wrong page.

What Logically gets right

It is a genuinely wide product, and breadth is a real feature when your alternative is five tabs. One login gives you a reference manager with folders, tags, over 10,000 citation styles, and imports from Zotero, Mendeley, and EndNote. It gives you a file annotator with highlights, sticky notes, and comments. It gives you a document writer. It gives you a Microsoft Word plugin for inserting citations where you actually write.

The AI research assistant is the part worth studying. It runs in three modes: retrieval over your uploaded documents, search across Semantic Scholar, and plain Google search, all returning inline citations with the data source tracked. That is a sensible design. When your own library does not contain the answer, the tool does not simply fail, it goes and looks.

It is also cheap. There is a real free tier with daily caps, and the paid individual plan sits in the low teens in USD depending on the listing and the billing term. For a student with a thesis and no budget, that matters more than any feature comparison.

Agent Bayes does not have a reference manager or a Word plugin, and it does not search the open literature. Those are three things Logically does that we do not.

Where the two tools separate

Depth of reading. Logically's support material suggests best results at roughly 20 to 30 connected documents. Agent Bayes is built for knowledge bases that grow past that, with plan limits measured in files per knowledge base and knowledge bases per project. More to the point, every PDF goes through OCR that preserves multicolumn text, tables, figure captions, footnotes and headers, then semantic chunking, then distillation into structured bullet points. Dense academic PDFs are the specific thing that pipeline exists for.

What the citation points at. Logically produces inline citations with data-source tracking. We could not find page-level citation into the source PDF documented as of 3 August 2026, so treat this as a difference in what each vendor publishes rather than a proven absence. On our side the model is explicit: author, year, printed page range, and the exact chunk of source text, opened in a built-in reader at that page. Printed page numbers, not PDF viewer positions, so they drop straight into a bibliography.

Where you read the paper. Both tools let you annotate PDFs. Ours is wired into the retrieval. Open any indexed document and the AI bullet points sit in a sidebar that keeps pace as you scroll, so the extracted summary of the passage in front of you is always the one on screen, and clicking a bullet jumps the reader to it. The sidebar also runs semantic search inside that one file, and collects your own highlights, icons, and comments, filterable by color, icon, or keyword. Colored page bookmarks sit in the toolbar for the pages you keep returning to.

The part that matters is what those bullets are. They are not a separate reading aid. Each one is the same citation item the retrieval engine returns and the agent attaches to a mindmap node, so a bullet you label while reading on page 14 turns up later in your labeled evidence collections, across every document in the project, ready to pin onto a claim. Reading, retrieval, and citation run on one set of objects rather than three.

What you are left holding. Logically leaves you a library, some annotated files, and a document. Agent Bayes leaves you a mindmap that keeps developing: conversation history per map, a rolling version stack for undo and redo, named snapshots, and provenance recorded on every change. You can replay exactly what the agent altered on the last pass, with additions, deletions, and word-level rewrites highlighted, before you accept it.

The corpus boundary. This is the trade in plain terms. Logically can leave your library. Agent Bayes cannot, by design. Their version is more flexible. Ours is more defensible, because when you can name every document that could have produced an answer, you can defend the answer.

What Agent Bayes does with a corpus

The agent works on your instruction through a pipeline that retrieves and ranks evidence, separates what a source argues from what it merely reports, writes cited claims into your mindmap, then scores the result against what you actually asked for and runs another pass if gaps remain.

When your sources conflict, the competing positions become sibling nodes with separate evidence chains rather than one smoothed statement, and you can ask for a cited prose discussion that moves through each school of thought.

The map is yours to edit. Rewrite any node, restructure branches, attach and remove citations, embed images, and pin nodes as context for the next instruction. The agent continues from the version you shaped, not from its own last output.

You can also audit your own writing. Select nodes and Agent Bayes reads the full text behind each attached citation, scores how well the evidence supports the claim as written, and tightens phrasing that overstates.

And there is a terminology graph, built from your documents rather than from a citation index. It extracts domain-specific terms from your corpus and connects them by how often they are discussed in the same passage, so you can search a term, explore its neighbourhood, and click any connection to read the passages behind it.

Alternative to Logically, or a deeper desk for one job?

Both readings are fair, and the choice comes down to where your time actually goes.

If the hard part is organising references and getting words into Word, Logically covers more of that in one place than we do, and it costs less.

If the hard part is working out what a stack of dense papers supports, where they disagree, and how to defend a claim to page level, that is the job Agent Bayes was built for, and it is the only job it does.