One of the fastest-rising repositories on GitHub right now is not a model or a framework but a memory layer for coding agents. claude-mem, built by developer Alex Newman (thedotmack), has climbed to roughly 96,000 to 98,000 stars, adding hundreds in a single day.
The pitch addresses a familiar frustration: an agent that solves a problem on Monday starts from zero on Tuesday. claude-mem hooks into an agent's sessions, captures what happened, compresses it into short observations using a model, stores them locally — in SQLite with a vector index for search — and injects the relevant pieces back into later sessions. It exposes its search through the Model Context Protocol, the interface a growing number of tools use to talk to agents.
It is deliberately agent-agnostic. The project lists support for Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot and OpenCode, and sells itself as 'persistent context across sessions for every agent'.
The repository is moving fast. A recent 'merge sweep' saw 54 community bug-fix pull requests merged in a single pass, alongside fixes for Windows port handling, semantic-search ranking and a bug that could send one worker into more than 2,000 restart cycles a day. Alongside the open-source tool, the maintainer runs a paid tier, CMEM Pro, that bundles a hosted 'observer' model and cloud sync.
claude-mem is not alone. It sits in a surge of projects treating memory as core agent infrastructure — from mem0 and projectmem to Cognition's Devin, which gained a readable long-term memory and a 'dreaming' process in early October. The common bet is that context windows alone will not be enough: agents that work over weeks need somewhere to keep what they learn.
