RECALL-Agent: Coding Agent with Long-Term Memory

Multi-agent coding assistant that learns from its mistakes: 75% task success with memory vs 0% without.

Planner, worker and critic agents work together: they plan a task, write code, test it in a sandbox, and store the lessons from each run in SQLite and ChromaDB so later tasks go better. I measured it with a reproducible benchmark (4 rounds, 3 seeds, 9 task families): 75% success with memory enabled, 0% without. Tools: Python, LangGraph, ChromaDB, SQLite, FastAPI, Streamlit; OpenAI, Anthropic and offline model providers.

Code: https://github.com/Shahidali78/recall-agent