Agent memory repo tops MCP charts on plain markdown

tigerless-labs/agent-memory reached 822 stars by September 10, built on a pitch that skips the vector database for agent memory entirely. It's the second markdown-and-git-native memory project to top GitHub's mcp topic in two weeks. That timing suggests vector databases are no longer the default choice for storing what an agent remembers.

Created September 1, the project stores long-term agent memory as plain Markdown files and treats that text as the source of truth.

Daily star counts ran 91, 103, 138, 125, 128, 173, then 64. Five of those seven days topped 100 stars, a stretch of sustained pull rather than one launch-day spike that fades.

Retrieval runs locally, ranking results against the markdown store with no external vector index involved.

A separate process, called the Manage layer, runs on a sleep-time schedule independent of the main agent loop. The repo doesn't say what work that layer performs.

Claude Code and Codex read from the same memory store, so one memory pool backs both tools instead of one store per agent.