← All verdict cards · Verify this card's timestamp ↗
By 2027-07-14, at least 4 of 5 major… — resolves 2027-07-14
By 2027-07-14, at least 4 of 5 major… — resolves 2027-07-14
The call
By 2027-07-14, at least 4 of 5 major coding agent products (Claude Code, Cursor, GitHub Copilot, Windsurf, OpenAI Codex CLI) document their primary memory mechanism as a long context window rather than fine-tuned weights or retrieval alone.
Sources
arXiv (Nelson Liu, Kevin Lin, Hewitt, Paranjape, Bevilacqua, Petroni, Percy Liang)
「we observe that performance is often highest when …」
The receipt
On 2027-07-14, read each vendor's official technical documentation or engineering blog and classify the primary memory architecture as context-window, weight-based, or retrieval-only. Four or more in the context-window class is a hit.
Resolves
Jul 14, 2027
Outcome
Not due yet — no public outcome to report.
What we learned
Not resolved yet — here is the condition we pre-committed to that would overturn this call (check it yourself, you don't have to wait):
If 2 or more shift to weight-based memory or pure retrieval as the primary mechanism, the long-context reading flips.
