Gemini’s “Thinking mode” just changed — and a lot of people are suddenly realizing how much they relied on hidden chain-of-thought and fuzzy “memory”.
For dev work, that hurts in three ways:
- You lose visible reasoning (how did it get that answer?)
- You lose reliable continuity (what does it still remember?)
- You have zero control over how/where any of that is stored
The QonQrete Approach
When I started building QonQrete, I took the opposite approach:
Reasoning and memory should live on your machine, in files you own — not inside a black box UI.
Instead of treating Gemini/ChatGPT/etc as “the place where the thinking lives”, QonQrete treats them as stateless LLM engines behind a local-first pipeline.
The Core Agents
- InstruQtor – plans the work (TasQ → briqs)
- ConstruQtor – applies the changes in an isolated repo (“qage” / qodeyard)
- InspeQtor – reviews everything and writes a reqap (retrospective/feedback)
Everything Goes to Disk
Each step writes its artifacts to disk:
- briqs = the full breakdown of the task (reasoning / plan)
- exeq summaries = what actually happened in the cycle
- reqaps = assessment + next steps
- qodeyard = the current code state
- struqture logs = raw agent output and events
That means:
- The “thinking” is not ephemeral UI — it’s markdown files and logs you can diff, grep, and version-control.
- “Memory” is not a mysterious cloud feature — it’s the accumulated TasQs, briqs, reqaps and repo state inside a qage.
How Continuity Works
- Cycle 1:
tasq.md → briqs → qodeyard → reqap - Cycle 2: the reqap is promoted to the next TasQ → new briqs → updated qodeyard → new reqap
- Cycle N: context keeps accumulating as files + code.
You can also resume from any “brain state”:
- drop files into
sqrapyard/ - put a previous
reqapintosqrapyard/tasq.md - start a new cyQle → QonQrete continues from there
Own Your AI’s Thinking
So while cloud UIs are debating how much “Thinking mode” you’re allowed to see, QonQrete just writes everything to disk and lets you own it:
- Chain-of-thought as files
- Memory as a repo + artifacts
- Full audit trail under your control
If you’re interested in local-first, file-based AI workflows (especially for code), the repo is here:
🔗 github.com/illdynamics/qonqrete
Curious what other engineers think: is this “Unix-style, file-based mindstack” the right direction, or are we all still too used to letting vendor chat UIs be our memory?