Persistent project memory for AI coding agents (Claude Code, Codex, Cursor) with architecture briefs, ADRs, and DevLogs.

nova

Give your AI coding agent a persistent memory across sessions.

中文文档 | Design Document | 设计文档

The Problem

AI coding agents start every session as a blank slate. They forget prior architecture decisions, undocumented debugging discoveries, and cross-project workflow lessons.

That causes repeated mistakes, decision drift, and poor continuity across both coding and non-coding sessions.

The Solution

nova gives Claude Code, Codex, and Cursor a layered memory system with both durable memory and snapshot memory:

Durable memories are loaded only when the task justifies them. Simple self-contained requests can skip session memory entirely. When the agent detects non-trivial, ambiguous, or history-sensitive work, it loads user.md, memory.md, and arch.md, then runs durable recall. Snapshot memories are synced out-of-band through ~/.nova/bin/nova-snapshot and can be auto-synced by an optional scheduler, but they are only searched when the user explicitly asks for historical conversations. Explicit "remember this" requests use a dedicated write path that recalls related memory before updating user.md, memory.md, or notes/.

How It Works

Session Start                           Background / On Demand               Code Commit
     │