Every Claude Code session starts the same way — re-explaining your stack, your conventions, what you already tried last week. You /clear or hit the context limit and everything is gone.

Grepping through JSONL session files works but it's slow, eats tokens, and gives you walls of text with no structure. You can't grep your way to "what uses Pydantic" across 50 sessions.

session-graph parses your session logs, extracts (s,p,o) triples with an LLM, links entities to Wikidata, and loads into Fuseki. One SPARQL query replaces 10 grep calls.

52 sessions → 138K triples, ~$0.70. Works with Claude Code, DeepSeek, Grok, Warp, etc.

https://github.com/robertoshimizu/session-graph