CITATION.cff

cff-version: 1.2.0
title: >-
  The Cost of Remembering: Filesystem Memory Against Long Context on
  LongMemEval
message: >-
  If you use this harness, the run data, or the paper, please cite it.
type: software
authors:
  - given-names: Jake
    family-names: Van Clief
    email: info@eduba.io
    affiliation: Eduba
  - given-names: David
    family-names: McDermott
    affiliation: Eduba
  - given-names: Kay
    family-names: Kumar
    affiliation: Eduba
license: MIT
year: 2026
abstract: >-
  Harness, run data, and paper source for an evaluation of filesystem-based
  agent memory on LongMemEval. An agent files each conversation into a
  directory of markdown files as it arrives, with no embedding model and no
  database; a second agent later answers questions by walking those folders.
  Accuracy is statistically indistinguishable from placing the whole history
  in a frontier model's context window, at 97% fewer tokens read and 95% lower
  cost per question.
keywords:
  - long-term memory
  - context engineering
  - retrieval
  - agent architecture
  - LongMemEval
references:
  - type: conference-paper
    title: >-
      LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive
      Memory
    authors:
      - family-names: Wu
        given-names: Di
      - family-names: Wang
        given-names: Hongwei
      - family-names: Yu
        given-names: Wenhao
      - family-names: Zhang
        given-names: Yuwei
      - family-names: Chang
        given-names: Kai-Wei
      - family-names: Yu
        given-names: Dong
    year: 2025
    url: https://arxiv.org/abs/2410.10813
  - type: article
    title: >-
      Interpretable Context Methodology: Folder Structure as Agent
      Architecture
    authors:
      - family-names: Van Clief
        given-names: Jake
      - family-names: McDermott
        given-names: David
    year: 2026
    url: https://arxiv.org/abs/2603.16021