CITATION.cff

cff-version: 1.2.0
message: "If you use Evensong / the Hybrid RaR benchmark harness in your work, please cite it."
title: "Evensong: Reverse-engineered Claude Code with Hybrid Retrieve-and-Rerank benchmarks"
authors:
  - family-names: "Zhu"
    given-names: "Hengyuan"
    alias: "Fearvox"
    email: "admin@zonicdesign.art"
type: software
repository-code: "https://github.com/Fearvox/Evensong"
url: "https://github.com/Fearvox/Evensong"
license: Apache-2.0
license-url: "https://github.com/Fearvox/Evensong/blob/main/LICENSE-APACHE"
version: "0.1.0"
date-released: "2026-04-19"
keywords:
  - retrieval-augmented-generation
  - hybrid-search
  - bm25
  - llm-reranking
  - agent-memory
  - benchmark
  - claude-code
  - typescript
  - bun
abstract: >-
  Evensong is a reverse-engineered, modifiable implementation of Anthropic's
  Claude Code CLI, together with a production-grade Hybrid Retrieve-and-Rerank
  (RaR) benchmark suite for agent memory systems. The benchmark uses a
  cross-LLM design (grok-3 generator + deepseek-v3.2 judge) over 648 trials
  (108 generated queries × 2 pipelines × 3 runs) on a 200-entry knowledge
  vault, showing a hybrid BM25 + LLM-rerank pipeline achieving +2.5pp top-1
  accuracy over an LLM-only baseline while cutting p50 latency by 27%, p90
  latency by 63%, and LLM prompt token cost by 75%.
preferred-citation:
  type: software
  title: "Evensong: Reverse-engineered Claude Code with Hybrid Retrieve-and-Rerank benchmarks"
  authors:
    - family-names: "Zhu"
      given-names: "Hengyuan"
      alias: "Fearvox"
      email: "admin@zonicdesign.art"
  year: 2026
  url: "https://github.com/Fearvox/Evensong"
  version: "0.1.0"
references:
  - type: article
    title: "EverMemOS: Memory for Self-Evolving AI Agents (arxiv 2601.02163)"
    authors:
      - name: "EverMind / Shanda"
    url: "https://arxiv.org/abs/2601.02163"
    notes: "Architectural inspiration for the Hybrid stage 1 + LLM rerank stage 2 design adopted in this repository."