hanishsaini/lore

Lore is a tool that gives AI coding assistants like Claude Code, Cursor, and Codex a memory of your team's past decisions. The author's framing is that these assistants know a lot about syntax, frameworks, and public APIs, but they know nothing about why your codebase looks the way it does. They will confidently suggest the same change that broke production last quarter because no one told them it was tried and reverted. Lore reads your git history, your architecture decision records, and your postmortems, pulls out the reasons behind past choices, and serves those reasons back to the assistant when it is about to make a related change. The headline example in the README shows the assistant about to add Redis caching to a payment endpoint. Before writing any code it queries Lore and gets back three relevant items: a previous Redis revert linked to a Stripe webhook race condition, a convention banning in-memory caching on payment writes, and an incident postmortem about duplicate charges. It then proposes a different approach. The pipeline behind that demo has a few stages. Lore ingests commits, pull requests, and markdown files for ADRs and postmortems. A signal pre-filter drops about 80 percent of routine work so the rest can be processed cheaply. A decision-extraction stage either calls a language model (Anthropic, Groq, Gemini, OpenRouter, DeepSeek, NVIDIA NIM, or local Ollama) or falls back to a pattern-based heuristic mode that needs no API key. Extracted decisions are embedded with a small local model and stored in SQLite using the sqlite-vec extension and FTS5 full-text search. Retrieval combines vector and keyword search through reciprocal rank fusion, then applies boosts for source type, severity, and recency. The assistant talks to Lore through MCP (Model Context Protocol), with tools named recall_decision, find_postmortems, and related_history. Everything is local-first; your data stays on your machine unless you wire it up to a cloud LLM yourself. Installation is pip install lore-cli, then lore init and lore ingest . inside a project. A lore demo command runs against a preloaded fictional payments codebase with no API key needed.