memory

A retrieval-augmented memory for local models. Before the model answers, this prompt-preprocessor retrieves the most relevant snippets from a folder of your notes/docs and injects them as context — grounding answers in your own knowledge base.

Part of lmstudio-suite.

How it works

On each message it embeds your query with an LM Studio embedding model, finds the top matching chunks from your knowledge directory (cosine similarity over a local vector index), and prepends them to the prompt. Indexing is cached and only re-runs when files change. Any failure passes your message through unchanged — retrieval never blocks you.

Configuration

Global:

• Embedding model — identifier of a loaded LM Studio embedding model (e.g. text-embedding-nomic-embed-text-v1.5).
• Knowledge directory — absolute path to a folder of .md/.txt files. Leave blank to disable.

Per-chat:

• Snippets to retrieve (top-K, default 4)
• Minimum similarity (0–1, default 0.35)
• Max injected context characters (default 2000)
• Enable memory write tools (off by default) — see below.

Writable memory (closing the loop)

Retrieval is read-only on its own. Turn on Enable memory write tools to also expose three tools so the model can save what it learns:

Tool	What it does
remember	Save a fact as a markdown note (frontmatter + tags) under memories/.
recall	Keyword-search saved memories and return matches with their ids.
forget	Delete a saved memory by id.

Because a remembered note is just a markdown file in the knowledge directory, the retrieval index rebuilds on the next message and the new fact becomes retrievable automatically — no separate store. (For structured, linked notes, use the kb-map plugin's write_node.)

Use

Point it at a notes folder and load an embedding model, then chat normally — relevant context is added automatically. Enable the write tools if you want the model to grow that folder over time.

MIT licensed.