Anyone who uses AI coding agents daily knows the ritual: open a new session, paste in your architecture decisions, remind the agent about naming conventions, re-explain your preferences. Every single time.
Bossa is a new open-source tool built to kill that loop. The concept is deliberately simple: it exposes a persistent filesystem to your AI agent through MCP (Model Context Protocol) or a standard CLI. Your agent gets ls, grep, read, and write commands over a storage layer that survives between sessions. Whatever context the agent writes during one conversation is still there in the next.
The interesting design choice here is what Bossa doesn't do. There are no vector embeddings, no retrieval-augmented generation (RAG) pipeline where documents get chunked and searched by semantic similarity, no fancy memory architecture. The creators argue that AI agents already know how to navigate filesystems, so there is nothing new for them to learn. The agent just reads and writes files the way it already does with your codebase.
This approach trades sophistication for predictability. You can inspect exactly what the agent stored, edit it by hand, and move it between machines by copying files. That transparency is a real advantage over black-box memory systems where you have no idea what the agent "remembers" or why it is behaving a certain way.
The MCP integration means Bossa slots into tools like Claude Code and Cursor without custom plugins. For developers already using MCP-compatible agents, adding persistent memory is a config change rather than a workflow overhaul.
The obvious limitation: filesystem memory works well for structured notes and preferences, but it will not scale to thousands of documents the way an embedding-based system would. For the target use case of personal context and project conventions, though, flat files are probably more than enough.