Features
Search by meaning
From desktop 1.99.110, Vector can also search your documents by what a request means, not only by its words: "how long do sessions last" finds "refresh tokens expire after 14 days". It does this with a small local model you download once. Until you do, Vector searches documents by keywords and code links, and everything else works the same.
The download
Nothing is downloaded until you ask, in Settings → Memory → Search model, from the offer on the Documents
tab once you have documents, or with vector memory model download in the terminal. Each place
shows the size, the license and the source first.
- Search by meaning
- Snowflake Arctic Embed XS, int8, 24 MB. Apache-2.0, by Snowflake, pinned to revision
d8c86521100dofSnowflake/snowflake-arctic-embed-xson Hugging Face. The default. - Full precision
- The same model's full-precision graph, 91 MB, offered only when the int8 one fails its check on your computer. Not yet verified: it has not been run in this release, and the check it must pass on your computer has not been measured on it.
- Code-aware (preview)
- IBM Granite Embedding Small English R2, int8, 55 MB, Apache-2.0, trained on code as well as prose. Offered on request; its check vectors are not pinned yet, so it is checked by how it ranks a set of test sentences.
Files come from Vector's own copy and then from Hugging Face at the pinned revision. A file is kept only at
its pinned size and sha256 (the default model's graph is e6aa5e656466a73d…), except the
Code-aware preview's tokenizer.json and tokenizer_config.json, whose hashes are not
pinned yet: they are checked by size and taken only from Hugging Face at the preview's pinned revision. A
cut download resumes where it stopped, and a redirect to a host the source does not use is refused. The
model lives in Vector's data folder, outside the brain's own folder, so erasing all memory keeps it. For a
computer with no network, vector memory model install --from <folder> takes the same files
from a folder, with the same checks.
After the download
The model runs on this computer, in Vector's background reader, with no network: requests and documents are
embedded locally and the vectors stay in brain.db. Each time the model loads, Vector embeds
three test sentences and compares them with the vectors the pinned model gave them; a model that gives
other numbers on your computer is not used, and Settings offers the full-precision one instead, which is not
yet verified.
- Records firstEach document's records are embedded before its passages, in the background, and its progress shows as Meaning on its row.
- Gentle on your computerAfter each batch Vector rests as long as the batch took, and three times as long while an agent is working.
- PausePause embedding in Settings → Memory → Search model, or with
vector memory model pause. What is embedded is still searched. - Never in the wayA request does not wait for a model that is still loading or slow to answer: it goes on with keywords and code links.
What it adds to a request
Search by meaning is one more way to reach a card, beside keywords and code links, inside the same budget described in Vector Brain. A card it finds has to stand out: every text looks a little alike to such a model, so Vector scores how far a passage is from what all text shares, and keeps only matches over a floor and close to the best one. A greeting has no meaning to search for and still carries nothing. Code itself is never embedded; code is reached through code links.
Private documents are searched by meaning only where they are searched at all: never for a free model labelled May train on your prompts, or in a session a guest is in. Removing the model deletes the vectors made with it.
In the terminal
vector memory model comes with the next CLI release, like the
rest of vector memory. A one-off command such as vector memory doc search loads
the model before it searches.
vector memory model status- Whether the search model is downloaded, how it was checked on this computer, and how much of this repository's documents it has embedded.
vector memory model download- Download it once, after it shows the size, license and source and asks; --yes skips the question, --code-aware takes the Code-aware preview, and --full-precision the full-precision model, which is not yet verified.
vector memory model install --from <folder>- Install it from a folder of its pinned files, for a computer with no network.
vector memory model pause- Hold embedding in the background; resume goes on with it.
vector memory model remove- Delete the model and the vectors made with it.
What the eval measured
With the search model, the same requests found the governing requirement for 13 of 40 tasks by keywords and meaning alone, with no code links, against 1 of 40 by keywords alone; with every channel at once, for 40 of 40, and 8 of 8 through an import. The 10 unrelated requests carried no tokens with it either.
The least score a match by meaning needs, 0.2, was set on this same eval. There the unrelated requests carried no cards from 0.18 up, and at 0.17 one of them carried 122 tokens, so the floor sits 0.02 above the least that kept them out, and these figures are not an out-of-sample result. The eval is five short specs for a small invented repository and 58 requests, written alongside the feature, with no language model call. It is not an independent benchmark. The dated report and every request it ran are in docs/benchmarks/vector-knowledge-meaning-2026-10-19 in Vector's repository.