thedotmack/claude-mem
Claude-Mem records agent activity, compresses it into persistent project memory, and retrieves selected history later, trading extra local services and stored data for cross-session continuity.
Claude-Mem captures lifecycle events, stores sessions and summaries in SQLite, and combines keyword and vector retrieval. Its **3-layer workflow** searches a compact index, inspects nearby timeline context, then fetches selected observations.
For recurring agent work, retrieve narrowly instead of injecting whole session histories. The project claims **about 10× token savings** from filtering before full fetches, and exposes privacy tags plus controls over injected context.
Claude-Mem captures lifecycle events, stores sessions and summaries in SQLite, and combines keyword and vector retrieval. Its **3-layer workflow** searches a compact index, inspects nearby timeline context, then fetches selected observations. For recurring agent work, retrieve narrowly instead of injecting whole session histories. The project claims **about 10× token savings** from filtering before full fetches, and exposes privacy tags plus controls over injected context. The plugin adds a Bun-managed worker, hooks, local storage, and Chroma-based search, so it expands both operational and privacy surface area. Install through its dedicated installer or plugin commands; **npm install -g** provides only the library and does not configure the memory system.
Claude-Mem turns selective episodic retrieval into an installable local workflow: index first, inspect nearby context, then fetch only chosen observations. It confirms the value of bounded retrieval while making the operational and privacy costs concrete. Unlike learned consolidation, it keeps memory external and inspectable, but its token-savings claim does not resolve provenance, contradiction handling, or production governance.