The Internet Time Capsule

AI that reads archived web pages as they were, not as history remembers them. See what people actually felt before hindsight rewrote the story.

Comment

Inspiration

Every time you ask an AI about a historical event, it gives you the settled, textbook version -- influenced by hindsight. We wondered: what if you could strip away everything history added after the fact and see what people were actually saying and feeling in the moment. The Internet Time Capsule was born from that question.

What it does

The Internet Time Capsule is a forensic AI tool that retrieves real archived web pages from the Wayback Machine and generates hindsight-free sentiment analysis that describes the mood, discourse, and public reaction as if filing a report from that exact moment in time.

Key features:

  • Single Snapshot Mode -- Analyze any archived page with AI-generated contemporaneous sentiment analysis and verbatim quotes
  • Hindsight Gap -- Side-by-side comparison of what a modern AI would say vs. what the page actually conveyed, exposing how time rewrites narratives
  • Timeline Comparison -- Track how sentiment shifted across multiple snapshots within a year
  • Topic Explorer -- Describe any historical moment in natural language and AI suggests real archived pages to explore

How we built it

  • Wayback Machine CDX API for resolving and retrieving archived snapshots
  • Browser Use to deploy an autonomous browser agent that visually navigates legacy DOMs in Chromium
  • Fetch.ai uagents for agent-to-agent orchestration, with a hosted agent on Agentverse using the chat protocol
  • GPT-4o for grounded sentiment analysis with strict no-hindsight prompting
  • Streamlit for the interactive frontend with custom CSS animations and a time-travel UI theme

Challenges we ran into

  • Browser Use's API had breaking changes in v0.12.6 -- we had to discover that it uses its own ChatOpenAI class, not LangChain's, and that Browser() takes params directly instead of a config object
  • Wayback Machine timeouts and rate limits required retry logic and aggressive caching
  • Getting GPT to suggest specific article URLs instead of generic homepages for the Topic Explorer required multiple rounds of prompt engineering and a validation-retry loop
  • Streamlit's synchronous execution model required a custom asyncio.new_event_loop() pattern for all async calls

What we learned

  • Hindsight bias is everywhere -- even in how we ask AI to summarize the past. Constraining an LLM to only use visible page evidence produces dramatically different results than a standard summary.
  • The gap between "what we remember" and "what was actually written" is the most compelling part of the product. People expected the iPhone launch to be universally praised -- the archived skepticism is genuinely surprising.

What's next

  • Screenshot capture of the archived page alongside the AI analysis
  • Collaborative mode where multiple users can explore the same timeline together
  • Expanded archive sources beyond the Wayback Machine

Built With

  • agentverse
  • beautiful-soup
  • browser-use
  • fetch.ai-uagents
  • langchain
  • openai-gpt-4o
  • playwright
  • pydantic
  • python
  • streamlit
  • wayback-machine-cdx-api
Share this project:

Updates