AGENT / MEMORY

Agent memory: notes, tool memory, and a searchable archive

Agent memory is any state a coding agent reloads after the context window ends—markdown notes, vendor “memory” folders, or MCP memory graphs. A separate archive keeps full masked chats so any agent can search past sessions. AIChatMemory is independent of Claude Code, Codex, and Cursor.

Project and repo notes

Teams often commit AGENTS.md, CLAUDE.md, or similar instruction files. Claude Code auto memory and Cursor rules play the same role: small, editable text the agent reads at startup. They capture conventions, not every line of every thread.

Tool-native memory

Claude Code maintains manual and auto memory under each project. Third-party MCP memory servers store entities and preferences in local databases. Both help the next session start smarter; neither replaces grep across six months of Codex JSONL on two Macs.

Working memory vs long-term recall

Inside one session, the model’s context window is working memory—it disappears when the chat ends. Long-term recall needs something durable: files on disk, a memory MCP, or uploaded transcripts.

What a searchable archive adds

AIChatMemory masks Claude Code, Codex, and Cursor chats on your Mac, uploads them to your workspace, and exposes MCP search, count, and session. Use it when the question is “what did we decide in that deploy thread?” rather than “what rule should the agent follow tomorrow?” Masking is not a guarantee that no sensitive text remains.

Download for Mac