Continuity agent

Turn fragmented AI conversations into a source-backed map of how ideas evolve, so you can audit, correct, and reuse your reasoning.

Comment

Inspiration

I am not a developer. For several months, I used ChatGPT as a thinking partner for strategy, technology, projects, and decisions.

Out of curiosity, I exported my conversation history and discovered a multi-million-word archive. That revealed a paradox:

The more thought I produced, the harder it became to use.

Search could recover words, but not the path behind an idea: when an assumption appeared, why a conclusion changed, or which promising branch had simply been forgotten.

When I received the OpenAI Build Week email, I saw an opportunity to turn that accidental discovery into a working agent.

ChatGPT's native memory supports personalization. Continuity Agent solves a different problem: helping users inspect and reuse the evolution of their own reasoning from original sources.

What It Does

Continuity Agent transforms fragmented AI conversations into a source-backed map of how ideas evolve.

For a selected topic, it reconstructs:

  • chronology and what was knowable at each date;
  • facts, hypotheses, predictions, and decisions;
  • contradictions and later outcomes;
  • active, dormant, rejected, confirmed, and unresolved branches.

Every important item links back to its source and distinguishes extracted evidence from a model-proposed interpretation. Users can inspect, accept, correct, reject, or reopen a branch.

The audit is not an endpoint. It helps users understand where an idea came from, follow how the reasoning evolved, and continue working on it when needed.

Users can then generate a compact context package containing only selected and validated information for a new AI conversation.

How I Built It

I used Codex as my implementation partner. I brought the original problem, product logic, usage observations, and privacy requirements. Codex helped transform them into a working prototype.

The project includes:

  • a deterministic local ingestion pipeline;
  • structured schemas for messages, claims, decisions, predictions, and outcomes;
  • an interactive reasoning graph;
  • source inspection and human review controls;
  • append-only correction history;
  • selective Markdown context packages;
  • PowerShell and Python tests;
  • a fully synthetic demonstration.

The main Windows workflow uses PowerShell and standalone HTML, CSS, and JavaScript. It requires no server, API key, Python installation, or external database.

Challenges

The hardest challenge was preserving epistemic status across time. Later evidence must not leak into a reconstruction of what was known earlier. Model interpretations must not silently become facts, and dormant branches must not be labelled rejected without evidence.

Privacy was equally important. The public repository and demonstration contain only synthetic data. Private archives remain local.

What I Learned

The opportunity is not another intelligence layer. It is the control layer around AI:

persistence, provenance, chronology, correction, and selective reuse.

When AI makes idea generation inexpensive, selecting, connecting, testing, and recovering those ideas becomes more important.

What Comes Next

The next step is model-assisted graph generation across arbitrary local archives while preserving human validation.

The same approach could later support team decision memory across meetings, notes, email, and chat.

Links

Built With

Share this project:

Updates