novasplace/csm

When you chat with an AI assistant, the assistant forgets everything the moment the conversation ends. Every new session starts from a blank slate. Cross-Session Memory, or CSM, is a system that gives an AI coding assistant a durable memory that persists across conversations. The assistant can remember decisions made in previous sessions, lessons learned from past mistakes, architectural choices, and context about the project it is working on. The project is not simply a database of saved notes. It is a pipeline with several stages. Raw events from tool calls and conversations flow into a work journal. That journal is compressed into structured summaries. Those summaries are analyzed for patterns and turned into candidate beliefs. Candidates then pass through safety checks before they are promoted to durable memories. The final stage assembles a short context brief that the assistant reads at the start of each new session to orient itself. Each stage is optional, and the pipeline degrades gracefully if you only use the basic save and search functions. The safety design is deliberate. Nothing gets permanently stored without passing a confidence threshold and checks for duplicates and contradictions. All maintenance operations default to a dry-run mode, meaning they show what would change without actually changing anything until you confirm. Advisory blocks that the assistant reads are labeled as previews, not authoritative facts, to prevent confident-sounding but wrong memories from driving bad decisions. The system is built on PostgreSQL with the pgvector extension for semantic search, or SQLite as a lighter alternative. It exposes 27 tools across groups covering memory save and search, governance and deduplication, belief scanning and promotion, agent journaling, and checkpointing. There are 728 tests in the suite. The project is aimed at developers building or extending AI coding assistants and agents that need continuity across sessions. Installation requires Node.js and either a PostgreSQL database with pgvector enabled or SQLite. The README describes the full pipeline, all tool groups, and configuration options for database connections and threshold profiles. The project is in an early state and the README notes that breaking changes are possible.