Brief, mostly signal

Rubi: a knowledge graph that cuts Claude Code re-reading

A local, zero-dependency MCP knowledge graph. The authors measured 37% fewer steps, but a 15% cost increase when memory is written after every answer.

Read the source at github.com Discussion

What the research found

Rubi is a local, zero-dependency MCP-based knowledge graph for Claude Code and other agents; its authors report 71% of cost went to re-reading context and about 37% savings in simulation, with a measured A/B result of 37% fewer steps but a 15% cost increase when memory is written after every answer.

Claims and how well the evidence supports them

Analysis of 284 Claude Code transcripts found 71% of cost was re-reading context 99%
Early compaction replay shows about 37% lower total cost 98%
Rubi runs locally with zero dependencies and is a standard MCP server 98%
Writing memory after every answer increased cost 15% in short sessions 98%

Evidence

github.com https://github.com/YURIBEDRAN/rubi
news.ycombinator.com https://news.ycombinator.com/item?id=50145380

Actionability rated medium.