ChatGPT for biotech: what this AI cannot keep track of

This is not an argument that ChatGPT is weak. It is very good, it can act in connected tools, and it remembers across conversations. The question for someone running a drug development program is narrower: what exactly does it remember, and can you point at it?

Take the strengths first, because they are real and they matter in this job. It will read a forty-page contract research organization report and give you a straight answer about the bit you care about. It will turn a messy set of call notes into a coherent summary. It will draft the difficult email to a partner. It will help you think through a term sheet position at eleven at night when there is nobody to think it through with. For a biotech lead with no analyst and no chief of staff, that is a genuine change in what one person can carry.

What its memory is, in OpenAI's own words

OpenAI's help documentation describes memory as something that "helps ChatGPT automatically remember useful context from your chats, files, and connected apps to personalize your experience, so you don't have to repeat yourself as often." Personalize is the operative word. The purpose stated there is to make responses fit you better, not to maintain a record you can rely on.

The same documentation is unusually candid about the limits, and the honest version of this comparison is just to quote it. On the memory summary: it "will not include everything that ChatGPT remembers based on your chats." On the sources panel that shows why a response was personalized: it may "not show every factor or source that shaped a response." On the underlying mechanism: memory is "a continually updated synthesis of context from your past chats, which may be broader than what can be shown as individual items in a summary."

OpenAI also documents what went wrong with the previous approach, which is the most useful sentence on the page for anyone evaluating this: the older saved-memories system "often became stale and relied on users to manually manage updates," and memories "could also contradict one another." And on correcting something: to fully remove a fact you need to "delete every source where it appears, including past chats, archived chats, files, the memory summary," plus disconnect any connected apps holding it.

Yes, it can act. That is not the gap

A comparison that claims ChatGPT only talks is out of date. Its apps documentation says apps "may be able to take actions, such as creating or updating information in a connected service," with a permission model that names four levels: always ask, ask on any change, ask on important actions, which is the default, and never ask. The default reads from apps automatically and asks before anything with meaningful effect outside ChatGPT. Sending mail, deleting content, changing sharing permissions and handling health or financial information are all named as important actions.

So the difference is not capability and not caution. It is that there is no file of yours in the middle.

What that costs on a real program

Here is where it bites in drug development specifically. A partner's diligence question list arrives with a response date. Twenty-two questions, three of which need numbers out of the toxicology package and one of which asks about a comparable deal someone quoted from memory in a meeting six weeks ago. Over the following two weeks the tox report date moves because a method transfer slipped, the confidentiality agreement's expiry gets closer than anyone noticed, and a second interested party goes quiet.

You can work every one of those problems in a chat window and get good help each time. What you cannot do is open the deal and see the current answer. The response date lives in one conversation. The revised tox date lives in another. The comparable that was quoted from memory was never checked against anything. When the board asks in three weeks what the state of play is, the reconstruction happens the same way it always did, out of your inbox, on a Sunday. And the specific risk with a synthesis rather than a record is not that it forgets, it is that a stale fact gets restated confidently in a context where you cannot see which source it came from.

There is also a limit worth naming rather than skirting: nothing on this site has verified anything about running regulated or patient-identifiable material through a general assistant, and no page here will tell you it is fine. That is a question for whoever owns data governance at your company.

The alternative is a file, not a better chat

The other shape is a single-seat system installed on the buyer's own machine, in plain files they own. It reads the mail and meeting transcripts they are already in, files each item to the right program or partner, moves the dates and figures a message changes, opens the follow-ups that change creates, and drafts the reply for approval. Nothing outbound goes without an explicit yes. The important part is unglamorous: the state of the deal is a file. You can open it, read it, correct it, and hand it to your board without reconstructing anything. Because it is plain text you own, it stays readable with no tool at all.

That is also why the platform choice matters, and it is covered in Claude for biotech and the AI that runs locally, alongside the practical question of Claude Code for biotech against off-the-shelf AI tools. If you would rather everything lived in one workspace, read Notion for biotech. The concept these all circle is set out in what an AI company brain is and why a biotech needs one, and the incumbents are laid out in the full alternatives comparison.