cowork-mem

You persist project context across sessions by using SQLite-based storage with FTS5 full-text search and automatic capture hooks. This stores decisions, file edits, and summaries, allowing the assistant to recall prior work and maintain continuity without repeated context gathering.

Semantic search for retrieving session decisions works by combining an FTS5 full-text index with a TF-IDF semantic search layer. This enables concept-level retrieval of stored observations, file edits, and session summaries organized by project and session.

You can automatically capture file edits and tool outputs using PostToolUse session hooks. Auto-capture filters avoid trivial noise, while SessionStart recalls previous summaries and PreCompact marks compaction points to maintain relevant project continuity.

To keep private observations out of search results in a shared project memory, wrap the content in a <private> tag. The tagged content is stored in the SQLite database but explicitly excluded from FTS5 full-text and semantic search retrieval.

The best way to manage old session memory entries is using CLI scripts. These scripts allow you to add observations, view timelines, compact old entries to save space, and export stored decisions, file edits, and tool outputs for external use.

Persistent session memory works without external database dependencies by using a local SQLite database. It provides FTS5 full-text search and a TF-IDF semantic search layer entirely locally, ensuring data privacy and offline accessibility for project workflows.