Give Claude, Codex, OpenClaw, and other MCP clients one source-cited memory layer. Agents search sources and propose updates; people review what becomes durable. Run locally from GitHub or use hosted personal and team spaces with governed remote MCP access.
I built NeuraKeep after watching capable agents repeat known mistakes and treat old chat transcripts as truth. The missing layer was not a bigger context window—it was memory with provenance, conflict detection, review, and failure recall. NeuraKeep makes durable memory inspectable: agents can search sources and propose updates, while humans retain control over what becomes trusted memory.