Show HN: We Gave AI Coding Agents Production Memory

Your AI coding agents ship code fast. Codela checks every deploy against how your app actually behaves in production, learns from the ones that broke it, and teaches your agents so they don’t repeat the mistake. See what Codela learns from your repo Every revert and production fix in your history is a mistake your agents could make again. Codela turns them into rules they follow. Paste a public repo. Codela reads its last year of reverts and production fixes and shows the lessons your coding agents would get. Public repos only, nothing is written anywhere. Read it step by step Tests pass. Production breaks. The agent never finds out. $ npm test ✓ 24 tests passed $ git push → deployed to production Tests pass Your coding agent writes code and runs the tests locally. They pass, so it ships. KeyError: 'sku' ×37 p95 120ms → 480ms Production breaks Then production shows what the tests couldn't: a slow page, a new error, too many calls to another service. Claude Code no memory of this The agent never finds out Someone on your team fixes it, but the agent never learns what went wrong, so it makes the same mistake again next time. .codela/lessons.md + Batch inventory calls: + one per cart, not per item read by your agents With Codela The lesson reaches the agent Codela turns what happened in production into lessons your agents read before their next code change, so the same mistake doesn't ship twice. Get started Up and running in minutes Sign in with GitHub Pick the repos Codela should watch. It reads code and deploys, comments on pull requests and adds lessons to your repo. It only ever writes its own lesson files, never your code. Connect your observability tool Pick SigNoz, Datadog, Sentry, New Relic or Grafana and paste a read-only key. Codela checks it works and matches your repos to their services. Keys are encrypted. Lessons go into your repo Right after you connect, Codela checks your last 30 days of deploys and a year of Git history, so you can see what it would have caught before you rely on it. From then on every deploy is checked automatically. Each lesson arrives in your repo as a pull request, and Claude Code, Cursor and Codex read it before they write code. Want to check every lesson first? Switch to review before merge in Settings. /plugin marketplace add Subomi1234/codela-plugin /plugin install codela@codela /codela:setup Memory layer Your AI agent reads the memory before it writes Every production problem becomes a lesson. Your agents read the lessons before they write code. In Claude Code ⎿ Codela: 3 production lessons for acme/checkout loaded > Show stock levels next to each item in the cart ● Before I write this: Codela has a production lesson for this repo. Calling get_stock() once per cart item made checkout 4× slower after deploy a1b2c3d. So I'll fetch every item's stock in a single get_stock_many() call, in batches of 50, and pass the results into the cart view. ⎿ Edited checkout/cart.py (+18 −3) On pull requests Codela: production lessons This PR looks like it repeats a mistake this repo already made in production. Calling inventory-service once per item made checkout 4× slower Where: checkout/cart.py line 42: get_stock() is called inside a loop over the cart items Learned from: p95 latency 410ms → 1.7s after deploy a1b2c3d Why: one checkout request turned into 60+ calls to inventory-service. Fetch all items in one call with get_stock_many() Keep batches at 50 items or fewer; the service rejects larger ones Only lessons learned from this repo's own production data are checked. Codela Your dashboard Lessons Codela learned from production, all in one place. See every deploy Codela checked, what broke and why, and which changes came from AI agents. Invite your team: everyone's agents share the same lessons. Deploys checked 3 got worse Lessons in your repo read by your agents Pull requests warned repeated a mistake Latest lessons highBatch inventory calls: one request per cart, not per itemAI-written · seen 3× mediumSample tracing at 10% when a second tracer runsAI-written mediumRetry refunds only with an idempotency keyseen 2× Activity 4f1c2e9 · Got worse · new error after deploy: KeyError: 'sku' ×37 b7d03a1 · Fine: no regression Works with your editor, your AI agent, and your APM In our tests Benchmarks We tested Codela on Codela Claude Code · 8 runs each · repeated the mistake Turning on Sentry made Codela 3× slower (p95 16ms → 54ms), so Codela wrote a lesson. Then we asked Claude Code to add Datadog. Without the lesson it made the same mistake 8 out of 8 times. With Codela, never. Limits the code doesn't show 3 scenarios · 20 runs in total · repeated the mistake Example: a stock service that refuses requests with more than 50 items. The code doesn't say so. Across three limits like this, Claude got it wrong 20 out of 20 times. With Codela's lessons, never. Questions teams ask before they start. Alerts tell a person that something broke. Codela works out which code change caused it and turns that into a lesson your coding agents read, so the agent doesn't write the same thing again. It uses the monitoring tool you already have; it doesn't replace it. No. When you connect, Codela checks your last 30 days of deploys and reads a year of your Git history (reverts and production fixes), so you see what already broke production, which of those changes came from AI agents, and the lessons it would have written. You can review each lesson before it goes into your repo. Codela comments on the pull request with the lesson and the production numbers behind it, so the mistake is caught in review instead of in production. You can, but almost nobody does it after every incident, and first you have to work out which change caused the problem. Codela does both automatically after every deploy, with the numbers to back it up. It only writes a lesson when it finds that the code change caused the problem, not just that the timing lined up. You can delete any lesson from your dashboard, and it's removed from your repo too. Never. The only thing it writes is its lesson files (.codela/lessons.md, plus a line in CLAUDE.md and a section in AGENTS.md), through pull requests. You can turn that off in Settings. The code changes in the repos you pick, and read-only data from your monitoring tool: speed, error rates, error messages and traces around each deploy. Keys are checked live and stored encrypted, and your lessons are never shown to anyone outside your workspace. Free for 1 repo and 20 pull request checks a month, no card needed. Pro is $7.99 per person per month for unlimited repos, pull request checks and your whole team. Cancel anytime.