Watch an AI Learn in 3 Minutes: A Walkthrough of the TrialSleuth AI Demo

🎬 What this is: I put together the demo video for our project TrialSleuth AI, an agent that reviews a new drug before its first human trial and learns from every past result. This post walks through the same demo step by step, so you can follow along or try it yourself.

🎥 Watch the video · 🚀 Try it live · 💻 Code on GitHub

🎥 Watch the demo first

🎯 The one thing the demo had to show

A demo video has a few minutes to make one point land. For TrialSleuth, the point was:

This AI remembers, and it gets better when you teach it something new.

Everything in the walkthrough is built around that idea: show the AI without memory, show it with memory, then teach it one new fact and watch the result change.

🧪 Step 1: Meet the drug

The demo uses a made-up cancer drug called HLV-772. In the app, a pharma team fills in a short form:

Composition: a KRAS inhibitor with an aniline group and a basic amine, cleared by the liver enzyme CYP3A4
Dose: 800 mg, rising to 1200 mg
Patients: includes people with moderate liver problems
Trial design: first in humans, 6 people dosed on the same day
Other medicines: no restrictions

To most people that looks like a normal plan. To an experienced safety reviewer, there are several warning signs hiding in it. The question is whether the AI can spot them.

😶 Step 2: Review with memory OFF

First, we switch memory off. Now TrialSleuth acts like a regular chatbot that knows nothing about past drugs.

The answer sounds reasonable but generic: monitor the liver, watch for interactions. There's no evidence, no specifics, and nothing a trial team could act on with confidence.

This is the "before" shot, and it's important. Without it, viewers can't see what memory adds.

🧠 Step 3: Review with memory ON

Now we switch memory on and run the exact same review.

This time the agent uses Hindsight, an open-source memory system for AI agents, to recall relevant facts before writing anything. The report changes completely:

🧪 Liver risk: patients with weak livers can get 2–3× more drug in their blood (HB-02)
💊 Interaction risk: a similar past drug had 2.5× higher levels with certain common medicines (PC-102). The fix was to ban those medicines first
👥 Dosing risk: a past trial that dosed everyone on the same morning went badly (PC-801). The fix was to dose one person first

On the right of the screen, the app shows exactly which memories were recalled. In the video, this is the moment to slow down and let viewers read. It's what makes the review trustworthy: you can check every claim.

🌱 Step 4: Teach it one new result

This is my favourite part of the demo.

In the sidebar there's a form called "Teach the agent a new result". We add one new lab finding: a sister drug, HLV-650, blocked a heart channel (hERG) and lengthened the heart's QT interval in a dog study. This gets saved to memory as PC-702.

Then we run the review again.

A brand-new ❤️ Heart row appears, citing PC-702 and asking for heart tests and ECG monitoring before dosing. Nobody changed the code or the prompt. The agent simply retained one new memory and used it.

🪞 Step 5: Ask what it learned

To close the demo, we click "What has the agent learned?". This uses Hindsight's reflect feature, which looks across everything in memory and sums up the lessons, like a safety expert's notes after years of trials.

⚠️ Throughout the demo, the app never says a drug is "safe". It flags known risks so human experts can fix the plan. That message is in the video too, because it matters.

🎥 Tips I'd share for demoing an AI product

Show the "before" first. The memory-OFF review makes the memory-ON review impressive.
Change one thing at a time. Same drug, same question, one new fact. Then the cause of the change is obvious.
Let the evidence breathe. Pause on the recalled memories so viewers can read them.
Keep it short. Two to five minutes is plenty if every second shows something new.
Say what it doesn't do. Being honest about limits builds trust.

🧩 How the memory works (in one paragraph)

TrialSleuth stores three kinds of memory in a Hindsight memory bank: facts about the human body, outcomes of past drugs (with the fix that worked), and summaries of its own past reviews. For each new drug it recalls what's relevant, the language model writes a report that cites the memory IDs, and every new result is retained for next time. The Hindsight docs explain each step, and Vectorize's guide What is agent memory? explains the bigger idea.