AI-assisted software development
Maintained Project Memory
Using ORGANIZER.md as an explicit, versioned knowledge base for AI-assisted software development.
I have been using AI assistants such as ChatGPT and Codex for software development for several months. They have become extremely valuable tools, but I also noticed a recurring limitation.
A new session usually starts with only a partial understanding of previous work. Even when the model has access to project files, important architectural decisions, discarded approaches and practical lessons learned over time are often missing. Recovering this information repeatedly costs both time and attention.
This observation led me to experiment with a different approach.
A project organizer instead of conversation history
Alongside the traditional AGENTS.md file, I introduced a persistent ORGANIZER.md (“project organizer”). At the beginning of every development session, Codex reads this file before starting any work.
Unlike a conversation log, the organizer contains only information that is expected to remain useful over time:
- architectural decisions;
- design constraints;
- proven procedures;
- implementation pitfalls;
- discarded hypotheses and the reasons they were abandoned.
It is not intended to replace project documentation, nor does it contain complete conversation histories. Instead, it acts as a compact engineering memory.
A hierarchical structure
The main organizer intentionally remains small. Whenever a project grows, it simply references more specialized organizers dedicated to individual components or subsystems.
This keeps the primary context compact while allowing detailed knowledge to remain organized and easy to navigate.
Controlled updates
The organizer is not updated after every interaction. New entries are added only when a session produces knowledge that is likely to improve future work:
- a decision that should remain stable;
- a recurring mistake worth avoiding;
- a reliable implementation pattern;
- an architectural constraint that should never be forgotten.
Personal information and unnecessary conversational details are intentionally excluded.
Because the organizer is an ordinary text file, it is completely transparent, version-controlled together with the project, editable by the developer, and easy to review or reorganize at any time.
Periodic maintenance
An important addition is a scheduled review process. Using a periodic cron task, Codex revisits the organizer and evaluates whether existing notes are still consistent with the current state of the project.
The review never changes architectural decisions automatically. Instead, it identifies entries that appear obsolete, contradicted by later developments, or worthy of confirmation, leaving the final decision entirely to the developer.
As a result, the organizer becomes more than persistent memory. It becomes maintained memory.
Practical benefits
In daily development this approach has produced several advantages:
- much faster context recovery at the beginning of new sessions;
- fewer repeated discussions about decisions already taken;
- lower risk of reintroducing previously rejected solutions;
- improved continuity across long-running projects;
- explicit and auditable project knowledge.
The organizer evolves together with the software instead of simply accumulating information.
A possible direction for AI development tools
This experience suggests that AI-assisted development could benefit from official support for explicit project memory.
Rather than relying only on hidden conversational context, development environments could provide dedicated tools for proposing new organizer entries, reviewing suggested updates, organizing project knowledge, and periodically validating existing information.
Such a system would keep long-term project knowledge transparent, versionable and fully under the developer’s control.
In my experience, this has been one of the most effective improvements for maintaining continuity during AI-assisted software development.
Source and discussion
The public ORGANIZER.md repository contains the demonstration files and maintained examples. The article is also available on LinkedIn for comments and discussion.