some improvements and next directions - feedback is appreciated
Hey Product Hunt!
We launched ChatRecall.dev last week to stop AI context from evaporating between sessions. Since then, we’ve realised the chat logs themselves hold the exact architectural blueprint developers and teams are constantly losing.
Whenever an agent builds software, it makes dozens of micro-decisions and cuts corners ("we'll fix this later") that immediately get lost once the prompt session ends.
Today, we’re restructuring ChatRecall around two major upgrades: Decisions and Autonomous Tasks.
1. Persistent Architectural Decisions
Auto-Extracted: As you and your team chat with AI, ChatRecall automatically extracts architectural choices, rejected alternatives, and tech selections.
Context Guardrails via MCP: When an agent attempts to pick a library or steer the architecture, it queries your decision log over MCP. If a decision was already made (e.g., "we use PostgreSQL, not Mongo"), the agent is automatically corrected.
Editable Anywhere: Update decisions directly in the dashboard or tell your AI during a session to update a decision node.
2. Autonomous Tech-Debt Kanban
Instant Task Extraction: Tell an agent "just hardcode this for now to unblock the demo", and ChatRecall automatically generates a follow-up task on your Kanban board.
Closed-Loop Agent Resolution: Tasks are exposed over MCP. When your AI agent has idle cycles (or when you trigger auto-fix), it can pick the task, execute the changes, document its reasoning, and close the ticket—taking the human out of the routine cleanup loop.
3. Streamlined Dashboard & MCP-First Focus
Cleaned-up UI focused on the decision tree and task pipeline.
Full MCP sync across your entire team so everyone’s agents stay on the same page.
The Loop: A missed decision triggers a task --> AI agent picks and resolves it --> the decision board updates --> clean codebase without micro-managing prompts.
It's live and free to explore at chatrecall.dev.
Would love your thoughts: how is your team currently tracking architectural decisions and prompt-created tech debt across AI sessions?
Replies
The closed loop is the part I would watch. I never built an approve button into mine, on the theory that a person rubber-stamping a monotonous run is not really a check.
What saved me once was two numbers disagreeing. A legal workflow of mine ran green, cost 285 credits, and produced a memo I would have approved on sight. The arithmetic found five issues at rank two or above while only two were marked unfavourable.
If a task closes itself, something countable should have to agree first.
chat-recall
@cmumulle I agree that a task closing itself can be wrong or scary but ideally ig the verification process is good enough you can filter or validate wrongfully closed tasks. I mean how many times it happened for people to close a task as it did not have good enough definition of done. So it happens for humans as well and the gate should be the verification and validation and not a human. I would really like some feedback if you have any and if chat recall works for you.
@darkkraft Fair on the definition of done. Humans close badly defined tasks all the time.
My chart audit counts in code and then stops. One run recorded eleven checks, nine blocking, two to revise, nothing clear, and it emitted the decision rule as text with no conditional step behind it. The rule is stated and never executed.
I left it there because I could not write a rule I trusted for the last nine. Extraction plus a stated rule is most of the value.
Based on my epxerience: mine live in one file the agent reads at the start of every session. Roughly thirty one-line rules, each pointing at a longer doc for the reasoning, and the file only grows by a new rule or a new routing line
The part that actually decides whether a decision holds is narrower than the file though. Approx. 5 of those rules have a test behind them that fails when someone breaks them: a naming convention, a set of claims we retired and can't let creep back into the copy. Those have never been violated. The other 25 get argued around whenever a prompt pushes hard enough, because a rule an agent consults is advisory and a red test isn't.
I'd personally stop at closed-loop auto-fix. Extracting the task, yes, that's the useful one
chat-recall
@siarheihamanovich I completely agree. The goal of this is to try and to make your own chat history the decision maker . And I really see the fit - hey I am building 5 projects - I have used betterauth for 4 of them the 5th most probably will be better auth.
I think that having this is just providing the ai best practices based on yours and your teams decisions. Hooks and actions could complement all of this.
Let me know if you believe chat recall can help you here I really hope it does.
I like the idea of keeping decisions and tech-debt tasks connected. It makes the whole process easier to follow.
chat-recall
@david_turner12 yes exactly. I mean decisions can differ but you will not change auth for every project right? I am happy to receive feedback and try to improve your experience. Let me know if you have any .
@darkkraft Exactly I think keeping those decisions visible will make it much easier to stay consistent. I’ll explore it more and share any feedback I notice!