Skip to main content
Relic is an early memory layer for AI-native teams. It ingests engineering history — pull requests, issues, reviews, commits — structures it in a temporal knowledge graph, and serves grounded answers and executable skills to coding agents through a Model Context Protocol (MCP) server.

Audience

Developer infrastructure / AI agent tooling. Runs locally or as a team service.

Role

Founding engineer / co-creator (with Abhinav and Zidan) — capture/ingestion pipeline, deterministic data mapping, and token optimization.
Coding agents like Claude Code, Cursor, and Gemini CLI can query organizational context and run verified team procedures with clear source citations.

​
Tech stack

  • Language & workspace: Python 3.12 on a uv workspace
  • Graph database: FalkorDB (self-hosted via Docker in dev)
  • Skill store: SQLite
  • MCP framework: FastMCP
  • Graph memory engine: Graphiti

​
Two halves, cleanly decoupled

Relic is built around a clean split between its two primary halves, which interact only through injected interfaces (e.g. RecallFn) wired at the CLI composition root.

Resilient by design

The serving layer stays functional even if the graph database or LLM network is unreachable — it never imports graph or OpenAI dependencies directly.

​
Key primitives

The graph memory engine — entity/edge extraction via OpenAI, vector embedding generation, and semantic recall.
A multi-tenant graph database hosting partitioned per-repository graphs, chosen for RediSearch full-text integration and multi-database support.
Exposes registered skills dynamically as tools and markdown resources, plus search and memory-recall capabilities.

​
Local commands

The project is driven by just:

Read the architecture deep dive

The decoupled skill registry, deterministic ingestion with token budgeting, and per-repo graph partitioning with cited provenance.
Last modified on June 22, 2026