Cook Memory: Finding the Right Context Across Conversations

Austin Worthington

Long-term memory · Technical report

Cook Memory: Finding the Right Context Across Conversations

The right context. Even conversations later.

Austin WorthingtonClient Acquisition.io

Published September 5, 2026 · 5 pages

Download PDF Cite this paper

Abstract

Cook Memory combines planned keyword searches, semantic retrieval, reciprocal rank fusion, cross-encoder reranking, and answer-oriented ordering. On the recorded LongMemEval-S evaluation, the final pipeline reached 98.5% Recall@5, 98.9% Recall@10, and 90.2 mean reciprocal rank on a 0-100 scale. A single keyword-query baseline recorded 83.6% Recall@5.

The measured improvement is 14.9 percentage points in top-five recall within Cook’s evaluation. The research asks a practical question: can the agent retrieve the earlier exchange that contains the evidence needed now, despite unrelated conversations around it?

Results

Questions with a relevant result in the top 5 · %Scale 0–100

Cook MemoryAnswer-bearing excerpts

98.5%
agentmemoryPublished · gold sessions

95.2%

Cook: 476 answerable questions, scored on answer-bearing excerpts. agentmemory: published 500-question session-retrieval evaluation. Different retrieval units and cohorts; reference comparison, not an identical-harness rerun. Retrieval is separate from final-answer accuracy.

agentmemory’s published methodology ↗

Finding the topic is only the beginning.

Combining keyword and semantic search, then ranking for the answer itself, lifted Cook’s top-five recall from 83.6% to 98.5% in the recorded evaluation.

Cite this paper

BibTeX for LaTeX and reference managers.

BibTeX

@techreport{worthington2026memory,
  title       = {{Cook Memory: Finding the Right Context Across Conversations}},
  author      = {Worthington, Austin},
  year        = {2026},
  month       = sep,
  institution = {Client Acquisition.io},
  type        = {Technical Report},
  url         = {https://trycook.ai/research/memory}
}