An open-source project has released a plugin that stores Claude Code session data in a single local file to solve inter-session memory loss despite 200K token context windows.

The GitHub repository for Claude Brain details a one-file memory engine called mind.mv2 that captures session context, decisions, bugs and solutions, auto-injecting at each session start with sub-millisecond search via a Rust core (https://github.com/memvid/claude-brain). Installation requires a one-time GitHub plugin setup followed by marketplace addition of memvid/claude-brain; the file begins at ~70KB and stays under 5MB after a year of use with full git integration for versioning. This directly counters the stateless reset problem illustrated in the repo's example of repeated auth bug discussions.

Claude Brain connects to the 2023 MemGPT paper that frames LLMs as operating systems needing explicit memory tiers (https://arxiv.org/abs/2310.08560) and LangChain's memory modules released that same year for chaining conversational state (https://blog.langchain.dev/langchain-memory/). Primary source coverage on the repo focuses on developer productivity gains but omits how the local-only design bypasses cloud API memory services from Anthropic and OpenAI while enabling brain transfer via scp or teammate handoff. Original documentation also understates the alignment with autonomous agent frameworks such as Auto-GPT that similarly pursue persistent cognitive state outside chat interfaces.

The project exemplifies a wider pattern of users constructing personal AI brains, turning foundation models into project-specific cognitive architectures that maintain evolving knowledge graphs without vendor lock-in or recurring inference costs.