ai-memory: Markdown Memory: Handing Off from Claude Code to Codex

LLM coding agents lose the context of previous work when a session ends. ai-memory collects lifecycle observation data from agents, organizes it into a wiki format, and delivers necessary information to the next session within a limited scope. When resuming work interrupted in Claude Code with OpenAI Codex, there is no need to re-explain the architecture or previously failed approaches.

Stored memories consist of plain Markdown files rather than vector databases. Located within a Git repository, they can be searched with grep, viewed in Obsidian, and backed up with rsync. The key feature is managing context at the file system level without complex infrastructure management.

It supports various coding agents such as Claude Code, Codex, Cursor, and Gemini CLI. Through each agent's MCP configuration and lifecycle hooks, it injects handoffs from previous work at the start of a session and summarizes relevant observations upon termination. The goal is not to support only specific agents, but to ensure continuity between different tools.

It supports Linux, macOS, and Windows (WSL2) environments, providing Docker images and native binaries. It works with major LLM providers such as Anthropic, OpenAI, and Gemini, as well as local embedding models like Ollama and vLLM. It is suitable for developers who want to reduce context loss due to agent switching or session fragmentation in long-term projects.