Apex Studio

Scar Tissue — The Coding Agent That Forgets What Doesn't Help Most coding agents never get smarter. They just accumulate: every reflection saved, every wrong rule kept, until memory is noise. Scar Tissue treats every lesson as a hypothesis that has to earn its place. Failures are distilled into lessons — then every epoch, each lesson is ablated: its tasks are re-run wit h it and without it. Lessons that hurt, never help, or never get selected are deleted, with the measured reason recorded. The result: held-out pass rate climbs 0% → 100% while memory shrinks from 20+ piled-up lessons to 8 proven ones — one per hidden house rule. 🔗 GitHub (full source, 24 tests, benchmark): https://github.com/coldrazer/Scar-Tissue 📄 Fetch.ai lesson schema: https://github.com/coldrazer/Scar-Tissue/blob/main/docs/fetch-schema.json 🧠 HOW IT WORKS • Every failure + its eventual fix becomes a candidate lesson, classified by failure type • Lessons live in vector memory (MongoDB Atlas $vectorSearch, or local JSON) and are retrieved per task • Only 2 global + 3 per-task lesson slots exist — a useless lesson crowds out a useful one • Ablation re-runs each lesson's most relevant tasks with vs. without it, averaged over repeated probes • Survivors are born "canary" and auto-publish after a passing held-out eval; other agents query them through a Fetch.ai uAgent • Ships with container sandboxing, API-key tenancy, background job runner, and a hard cost budget ⚠️ HONEST NOTE The benchmark numbers in this video are from the offline mock model (stated noise rates, reproducible with `scar.py compare --mock --epochs 4 --seeds 3` in the repo). Real-model co mparisons (Meta Llama, Google Gemini, Subquadratic SubQ) run with the same command — the repo README shows how. 🛠️ BUILT ON The Meta ecosphere — Meta Wonder · Meta Spark · Muse Glimmer Research: Subquadratic (SubQ) · Front-end schema: Fetch.ai · Benchmark testing: Google Gemini Solver/distiller: Meta Llama API · Memory: MongoDB Atlas Built for the AI Engineer Code Summit hackathon — "systems that get measurably better with every run."