Claude Code memory: MemSearch, MemPalace and OpenBrain

This comparison organizes Claude Code memory into six levels, based on where information lives and how the coding assistant retrieves it. The speaker recommends starting with CLAUDE.md instructions and auto memory, keeping indexes short and putting detailed context in separate files.

The practical setup covers structured Markdown memory and a hook that injects project and global indexes before the first tool call. The Memsearch walkthrough then adds semantic retrieval, with a prompt-submit hook that supplies the top three matching memories. The speaker contrasts this with Claude Mem's MCP search tools and management features.

For exact conversation recall, the video describes MemPalace as a local system that stores verbatim text with SQL and Chroma databases. Its limitation here is readability: the stored conversations are not plain Markdown files. These local storage choices do not establish that Claude itself runs offline.

The later sections address different needs. Karpathy's LLM Wiki pattern separates source documents from a Claude-maintained Markdown wiki; Recall is presented as a hosted alternative. OpenBrain uses a Supabase-hosted Postgres database and MCP access to share memory across tools. The speaker notes its longer setup and external-query latency, then compares it with Mem0. His recommendation is to add complexity when missed context or work across multiple apps warrants it.