CLAUDE CODE / MEMORY

Claude Code memory, and what a separate archive adds

Claude Code memory keeps instructions and learned notes on your Mac—mostly markdown files Claude loads at session start. A separate archive keeps whole masked chats so any agent can search past work. AIChatMemory is independent of Anthropic; both can coexist.

Manual memory (CLAUDE.md)

You can maintain CLAUDE.md at the repo root, a global file under ~/.claude/CLAUDE.md, or a local-only CLAUDE.local.md. These files tell Claude how to build, test, and behave in a project. They are small, editable, and meant for rules—not full transcripts.

Auto memory

Claude Code can also write auto memory under each project’s memory/ folder when it learns conventions from your corrections. The index loads at startup; it summarizes habits, not every line of every past chat.

What memory does not replace

Memory files do not store complete conversation history. They will not answer “show me the deploy failure thread from March on my laptop at home” unless you already distilled that story into a note. Resume commands only reload one local JSONL session at a time.

What a separate archive adds

AIChatMemory uploads masked Claude Code, Codex, and Cursor chats from every Mac in your workspace. The MCP exposes search (meaning plus keywords), count (exact totals), and session (one masked chat in parts). Agents retrieve; they write the answer themselves—there is no server-side ask. Chats are masked on your Mac before upload; reads of .env and credential files are dropped. Masking is not a guarantee that no sensitive text remains.

Memory: compact rules Claude should follow next time.
Local history: full threads on one Mac, opened with resume.
Archive: masked chats from every Mac, searchable by any connected agent.

Download for Mac