OptMem hits 1,000 GitHub stars with zero infrastructure

VictorTaelin's OptMem crossed 1,000 GitHub stars six days after its July 25 launch by skipping the usual memory stack. For anyone deciding whether an agent needs a memory database, OptMem claims a million memories fit in 608 MB and retrieve in 0.03 seconds.

No vector database, no embedding model, no background process: every memory is a line under 280 bytes, written to LOG.txt as it happens. Retrieval runs through a binary tree of one-line summaries: #0-1 and #2-3 fold into #0-3, and so on up the file. Those merges aren't automatic. They run on demand when you call memo nap, keeping the summary tree a rebuildable cache rather than a live index.

Installing it means pasting a 426-token markdown block into an AGENTS.md or CLAUDE.md file. There's no separate setup step beyond that. The project's own documentation flags one gap: subagents can't use memo independently, so multi-subagent setups aren't covered yet.

What's untested outside VictorTaelin's own repo is whether the binary-tree-of-summaries approach holds up on messy, non-uniform memories instead of tidy log lines. It's also unclear if the 0.03-second number holds at real scale, not just in a demo.