The design choice worth knowing about is deletion: consolidation folds several near-duplicate facts into one summary and archives the originals, and decay lets unused memories fade. That is the opposite of a vector store that keeps everything, and it is what stops recall from drowning in old noise.
A memory store that runs on your machine. Facts are extracted, deduplicated, embedded and linked by shared entities; related ones are later compressed into a single summary, and stale ones decay on a forgetting curve.
- Recall ranked by similarity and by how strong a memory still is, with a boost for memories tagged to the directory you are working in
- Storing a new fact with an importance value, a category, and optional file or directory paths
- Updating a stored fact when it goes out of date - re-embedded, and the change written to the audit trail
- A hash-chained ledger of every read, write and delete, so an altered record breaks the chain
- Shared team memory pools with roles, while private memories stay private
- One-command export of everything held, and a purge for the right to be forgotten
Python 3.11-3.14. Install with pip install yourmemory, register a token from yourmemoryai.xyz, then run yourmemory-setup, which wires up Claude Code, Claude Desktop, Cursor, Windsurf and Cline. Storage is DuckDB in a local file by default; Postgres with pgvector is the shared option. The licence is CC BY-NC 4.0 - noncommercial use only.
One command plus a key — pip install yourmemory, then supply credentials