akitaonrails/ai-memory: Git-Backed Long-Term Memory and Cross-Agent Handoffs
A technical review of akitaonrails/ai-memory, exploring cross-agent handoffs, Git-backed markdown storage, zero-LLM memory retention, and self-hosting constrain
Persisting contextual state across artificial intelligence coding harnesses has emerged as a fundamental workflow hurdle. According to official documentation, akitaonrails/ai-memory provides long-term memory for coding CLIs and facilitates structured handoffs across agent vendors.
Cross-Agent Handoffs and Shared Workstreams
As documented in the repository, the project connects over twenty harnesses, including Claude Code, OpenAI Codex, Cursor, and Devin. Developers can quit one agent mid-task and resume in another without re-explaining architectural context or failed attempts. Using ai-memory run transfers a shared workstream and a private, temporary local projection of native history. Handoffs use a typed protocol claimed exactly once, alongside cross-project user profiles.
Plain Markdown Wiki and Zero-LLM Retention
The documentation states the canonical source of truth is a Git-backed wiki of plain markdown files, with the database acting as a derived rebuildable index. The default workflow uses zero LLM calls: lifecycle hooks capture observations silently, and search works without API keys. Memory decay follows an access-weighted schedule, compacting cold notes into durable facts like error codes and paths. Edits remain recoverable via restore-page.
Self-Hosting Architecture, Privacy, and Constraints
Shipped as a self-contained binary for local or server hosting, it includes multi-user auth and an audit log. A typed privacy boundary ensures remote servers receive harness and session identity without local transcripts. However, repository limits apply: while Linux, macOS, and WSL2 are supported, Native Windows is experimental. Enabling the web UI also requires compiling with --features management-ui and running serve --transport http --enable-management-ui.