I stopped using ChatGPT's memory as project state and turned Google Drive into an external operational memory

With the help of Chatty (that’s what I call ChatGPT), I built a simple system for managing long-term projects without depending too much on ChatGPT’s built-in memory. The problem was pretty simple: ChatGPT was good at remembering things like how I prefer to work, but project information eventually became outdated. “I like to discuss the architecture before writing code” is useful long-term memory. “Version 1.2 has three bugs and this is the next task” is not. That’s project state, and project state changes all the time. So Chatty and I separated them. ChatGPT memory is mainly for stable things: preferences, methodology, general interests and the long-term identity of a project. Google Drive is now the operational memory: current project state, checkpoints, decisions, reusable skills, tests and important incidents. The rule we use when information conflicts is very simple: current file/source > AI_Workspace in Drive > ChatGPT memory > inference. In other words, old memory should never override a newer project file. I originally considered Obsidian, databases and more complicated setups, but realized I didn’t really need them yet. Google Drive was already available from ChatGPT, so we created an AI_Workspace folder there. The structure is basically: 00_System, 01_Projects, 02_Skills, 03_Checkpoints, 04_Decisions, 05_Tests, 06_Incidents and 07_Archive, plus an INDEX.md file at the root. We also tested whether ChatGPT could update the same Markdown file instead of constantly creating copies. It worked. The same Drive file ID was preserved while the content changed, which means a project can simply have something like STATE.md that evolves over time instead of STATE_final_v2_REAL.md forever. If someone wants to try something similar, this is basically how we did it: In ChatGPT go to Settings → Apps, find Google Drive and connect the Google account you want to use. Review the permissions and authorize it. Depending on your ChatGPT plan/workspace, the Drive actions available to you may vary, especially actions that create or modify files. Start a new conversation and tell ChatGPT that you want Google Drive to become your operational project memory. Ask it to create the workspace and test that it can create, read and update Markdown files. Add a short rule to Custom Instructions so this behavior is still there when you start a new chat. This was the setup prompt I used, adapted slightly so other people can copy it: *******************************************************************************************************I want to use Google Drive as an external operational memory for long-term projects. Create a folder in my Google Drive called AI_Workspace with this structure: AI_Workspace/ INDEX.md 00_System/ 01_Projects/ 02_Skills/ 03_Checkpoints/ 04_Decisions/ 05_Tests/ 06_Incidents/ 07_Archive/ Inside 00_System create: README_AI_Workspace.md Memory_policy.md Working_methodology.md Stable_memory.md Also create a Templates folder containing templates for: Project Checkpoint Decision Skill Test Incident The purpose of this system is to separate stable ChatGPT memory from changing project state. ChatGPT memory should mainly contain stable preferences, working methodology, general interests and long-term project identity. AI_Workspace should contain project state, checkpoints, decisions, skills, tests, incidents, pending work and other changing operational information. Use this authority hierarchy: current source or file > AI_Workspace > ChatGPT memory > inference. Before considering the setup complete, create a Markdown test file in Drive, read it back, update its content in place and verify that the same Google Drive file ID is preserved. Do not create unnecessary complexity. Keep everything readable in plain Markdown. ******************************************************************************************************* Then I added this to my ChatGPT Custom Instructions: *******************************************************************************************************Always speak to me in my preferred language. I use AI_Workspace in Google Drive as my canonical operational memory. When a request refers to an existing project and the current state is not sufficiently clear from the conversation, consult AI_Workspace before answering or reconstructing the state from historical memory. Recommended retrieval path: INDEX.md → relevant project → current checkpoint/STATE → relevant decisions → skills/tests/incidents if needed. ChatGPT memory should mainly be used for stable preferences, methodology, long-term project identity and general context. Operational information such as current state, pending tasks, versions, temporary decisions, errors, checkpoints, tests and incidents should live in AI_Workspace and should not be unnecessarily duplicated in memory. Authority hierarchy: current source or file > AI_Workspace > ChatGPT memory > inference. If AI_Workspace is unavailable or does not contain enough information to reconstruct the current project state, say so explicitly instead of inventing the missing state. Only update AI_Workspace when something operationally meaningful changes, such as a decision, progress, pending task, error, checkpoint or project state change. Do not turn every exploratory conversation into permanent project state. Do not consult Drive unnecessarily for casual questions, general knowledge or unrelated topics. ******************************************************************************************************* So now, if I start a new conversation and say “let’s continue Project X,” the idea is that Chatty first checks whether the current conversation already contains enough information. If it doesn’t, it goes to Drive, finds the current project state and continues from there instead of guessing from some old memory. If I ask something unrelated like “what is quantum computing?”, there’s no reason to touch Drive at all. One other thing we added was the idea of checkpoints and decisions. A checkpoint is basically a save game for a long AI collaboration. Decisions can also store why something was chosen and why alternatives were rejected. That way, six months later, neither the human nor the AI accidentally revives an idea that was already tested and discarded. We also use a simple principle of deterministic before AI. If something can be reliably solved with SQL, a script, a rule or a validator, we prefer that. The LLM is used where interpretation, reasoning, synthesis or ambiguity actually matters. The setup is still deliberately simple. No vector database, no custom agent framework, no complicated RAG stack and no special memory service. Right now it’s basically ChatGPT + Google Drive + Markdown + some discipline. The interesting part for me is that I started this thinking I needed to make ChatGPT remember more. I ended up doing almost the opposite: make it remember less, but make sure it knows where to retrieve the right information when it needs it. Has anyone here built something similar? I’m especially interested in hearing from people who have used an external-memory setup for months. What starts breaking after a while? What would you change?