Xerj open-sources Corpus Hub: agent reference memory that cuts token cost 57%
Roger_M_Taylor · x · 2026-10-14
Xerj launched Corpus Hub (Apache 2.0), a "reference memory" for AI agents: 515 open knowledge corpora across 11 domains (78.6 GB pinned text) that agents search instead of dumping entire files into context.
Same-task comparison: without Xerj, output tokens were 3× higher, cost $4.57 with 7/16 tasks completed; with Xerj, cost dropped to $1.98 (−57%) and 16/16 tasks completed. Runs locally, works with any model, covering domains from SEC and security exploits to crypto, law and health.
Related event: XERJ Corpus Hub Gives AI Agents Searchable Expert Memory(3 posts)→
More from coding & agent
- Production AI Systems Should Combine Multiple Knowledge Base Approaches — goyalshaliniuk · 2026-10-14
- Beyond RAG: Structured Data and Hybrid Search for AI Knowledge Bases — goyalshaliniuk · 2026-10-14
- Vector Databases vs Knowledge Graphs: Picking the Right AI Knowledge Base — goyalshaliniuk · 2026-10-14
- 7 Approaches to Building AI Knowledge Bases, from Vectors to Live Updates — goyalshaliniuk · 2026-10-14
- 7 approaches to building AI knowledge bases, from vector RAG to continuously updated pipelines — goyalshaliniuk · 2026-10-14
- Cursor User Burns $378 in One Night as Claude API Calls Blow Past $200 Ultra Plan — lxfater · 2026-10-14