The Excavatorium

Conversation text is sent to GPT-5.6 only when excavation starts.
Dashboard for active tools, buried tools, decisions, and open loops.
Export or restore the full archive as portable JSON.
GPT-5.6 extracts findings, decisions, open loops, and reusable prompts.
Reviewed excavations become permanent, linked conversation records.
Search across tools, conversations, decisions, tags, and record content.

Why I built it

I kept having long AI conversations that contained useful decisions, technical lessons, and reusable reasoning, only for all of it to disappear into chat history and that was honestly terribly annoying because I forgot all the useful things I've discussed with AI and I needed a way to retain useful decisions and judgments.

The Excavatorium is my private archive for preserving that judgment.

It stores four kinds of records:

Tools
Repositories
Conversations
Decisions

A tool record is more than a bookmark. It records what caught my attention, what the tool promised, what actually happened, what worked, what failed, the final verdict, and what would make me reconsider it.

Tools that fail badly enough can even be buried. Grok currently resides in the graveyard under the classification “Grok-tier cursed.” The story writes itself...

Conversation Excavation

During OpenAI Build Week, I added Conversation Excavation with GPT-5.6.

A user can paste a messy AI conversation and generate an editable structured draft containing:

a title and summary
high-signal findings
decisions made
open loops
reusable prompts
memory candidates
suggested links to existing records

The feature is not designed to make AI authoritative rather to use AI as a helper to compress and archive important decisions or ideas you're working on.

GPT-5.6 proposes the draft. The user reviews and edits it. Nothing enters the archive until the user applies the draft and uses the normal Save workflow as to preserve agency.

The result is not merely a summary. A temporary conversation becomes a durable judgment record connected to the rest of the archive.

How Codex was used

Codex was used throughout the repository workflow for inspection, bounded implementation, debugging, validation, and review.

It helped implement and harden Conversation Excavation, including work around:

OpenAI Responses API parsing
runtime validation
stale-response protection
candidate-record scope
privacy controls
authenticated Supabase Edge Function handling
restricted CORS
review-before-save behavior

Codex also helped verify the actual repository changes rather than relying only on generated output. I couldn't have done it without Codex and the new GPT 5.6 model which helped me brainstorm the app first of all and then helped me write it and improve my technical abilities. So I'd say Codex was instrumental in building this app and in maintaining it.

How GPT-5.6 was used

GPT-5.6 performs the conversation excavation itself and it was the catalyst for brainstorming and creating the app idea then the implementation.

It receives the conversation only after the user explicitly starts an excavation and returns a structured, reviewable proposal. The model does not directly write records into the archive.

Thus I'd say none of this would have been possible without the reasoning of 5.6 because I had a fuzzy idea of what I wanted and it helped crystallize the idea and implement it, especially it helped me learn coding basics etc.

Portability and privacy

Personally privacy is important to me so I had it in mind while creating this app, I made the explicit markdown and exports and json backups as ways to store my information and as further fuel for the app and further conversations with AI.

For example, I export a decision or tool conclusion I have made to ChatGPT and then discuss with it and stress test my ideas and conclusions.

The app is a knowledge base for longtime use as it only gains use and function the longer the user uses it as the knowledge and judgments accumulate. Also, the user can always inspect the decisions he/she has made to establish what made them abandon a tool or what made them stick with a certain repo/AI etc.

Individual records can be exported as readable Markdown for documentation, long-term storage, or use in later AI conversations.

The full archive can be exported and restored as a JSON backup containing the records, links, IDs, timestamps, and relationships.

The application uses a private Supabase database. Authentication and archive storage are separated from the public application shell.

The Excavatorium is fully responsive and works on both iPhone and desktop, allowing the archive to remain accessible and usable across devices.

Technology

The Excavatorium is built with React and TypeScript, hosted through Lovable, and backed by a directly managed Supabase project for authentication, PostgreSQL storage, and the authenticated Edge Function that calls GPT-5.6.

The central idea is simple:

AI conversations should not remain disposable.

The Excavatorium turns them into a living archive of judgment.