@charliermarsh: Some further thoughts on this... Every change was closely human-reviewed. That's the bar we have for ty -- we still do …
Summary
Charlie Marsh reports that a campaign with 5.6 Sol reduced ty's retained memory by 38% across ecosystem projects while improving performance, emphasizing that every change was closely human-reviewed.
View Cached Full Text
Cached at: 07/11/26, 11:29 PM
Some further thoughts on this…
Every change was closely human-reviewed. That’s the bar we have for ty – we still do very close human review, it’s an extremely complex project. Review is absolutely the bottleneck. I wonder when it will no longer feel required.
Charlie Marsh (@charliermarsh): We ran a campaign with 5.6 Sol to reduce ty’s retained memory and cut it by 38% across our ecosystem projects, all while improving performance. Completely insane. This is an enormous reduction!
In honor of the Bun blog post, I asked Codex to make a a Bun-style splash image.
Similar Articles
@charliermarsh: Since my last post, I reduced ty’s retained memory by another 15% with Codex. We're now at a ~25% memory reduction over…
Developer @charliermarsh reports reducing ty's retained memory by another 15% using Codex, achieving a total ~25% memory reduction via /goal, and expresses enthusiasm for GPT models.
@charliermarsh: It is incredible that with /goal you can just be constantly improving your software in the background. Make it faster, …
Charlie Marsh praises the /goal tool for enabling continuous background improvement of software, making it faster, smaller, and more stable.
@garrytan: This has sort of been the story of how gbrain has gotten better. As I use it for my 100K+ personal and company brains, …
Garry Tan shares ongoing improvements to gbrain, a retrieval tool for personal and company brains, noting weekly advancements.
@tricalt: https://x.com/tricalt/status/2057173322924806651
A founder discusses the scaling challenges of using markdown files for AI agent memory in production, highlighting common pitfalls with permissions, multi-agent interaction, and temporal queries, and suggests that teams often end up patching around these issues without realizing they are rebuilding a more complex system.
@hasantoxr: I'm replacing every memory layer I've ever built into an agent with this. SureThing dropped SOTA on LongMemEval. 88.0% …
SureThing has achieved state-of-the-art results on the LongMemEval benchmark, scoring 88.0% overall, prompting developers to replace existing memory layers in their AI agents.