I kept pasting the same context into Cursor and Claude Code every time I switched machines — architecture decisions, naming conventions, personal preferences. Tedious enough that I built something to fix it.
The idea is simple: give your agent ls, grep, read, and write over a persistent filesystem. Whatever it stores in one session is there in the next. No embeddings, no retrieval pipeline — agents already know how to navigate a filesystem, so there's nothing new to learn.
This also solves another problem. LLMs reason worse when context is bloated. A filesystem enables progressive disclosure naturally: the agent starts lean, runs ls to find what's available, greps for specifics, reads only what it needs. By the time it's reasoning about your actual problem, the context window is full of signal.
I tried RAG first. It doesn't map well to exact preferences — "use camelCase" shouldn't compete with semantically similar documents for retrieval rank.
Under the hood: Postgres with trigram full-text search. Each file is a row, directories are implicit from paths. Grep across 10K files in ~50ms.
To try it (sign up takes 30 seconds): pip install bossa-memory bossa signup bossa login bossa files write /memory/test.md "Hello" bossa files read /memory/test.md
Add the MCP server URL or CLI tool to Claude, Cursor, or LangChain and your agent gets ls/read/write/grep/glob.
Free tier is 100MB and 1K requests/day. Self-hosting is documented. Code is MIT.
Docs: https://docs.bossamemory.com GitHub: https://github.com/vinny380/bossa
Curious what memory patterns others have settled on and whether a plain filesystem abstraction fits how you think about it.