FAQ

Does memsearch work on Windows?

Yes, but Milvus Lite (the default local .db backend) does not provide Windows binaries.

If you are on Windows, use one of these options instead:

• Milvus Server via Docker
• Zilliz Cloud
• WSL2 if you specifically want the Milvus Lite local-file workflow

See Getting Started — Milvus Backends for the backend comparison and recommended setup.

How do I wipe and rebuild the index?

Use memsearch reset --yes to drop the current collection, then run memsearch index again.

memsearch reset --yes
memsearch index .

This deletes the indexed chunks in Milvus, but it does not delete your source markdown files.

For command details, see the CLI reference.

Why do I see "collection is in state 'released'" with Milvus Lite?

Milvus requires a collection to be loaded before search, query, or get operations. Some newer Milvus Lite / PyMilvus combinations reopen an existing local collection in a released state, especially across separate CLI invocations such as:

memsearch index .
memsearch search "my query"

Current memsearch versions explicitly load an existing collection before search/query operations. If you still see this error, upgrade memsearch first.

If the error started after upgrading Milvus Lite itself, also check whether you are reusing a local .db file created by an older Milvus Lite release. Milvus Lite 3.x was rewritten with a new pure-Python storage engine, and old .db files from the previous storage format are not compatible with the new engine. In that case, move the old .db file aside and rebuild from your source markdown files:

mv ~/.memsearch/milvus.db ~/.memsearch/milvus.db.bak
memsearch index . --force

If you do not want to rebuild the local database, keep using the older Milvus Lite environment that created it. For a more stable shared or long-running backend, use Milvus Server via Docker or Zilliz Cloud.

How do I see what is indexed?

Start with index stats:

memsearch stats

That shows the total indexed chunk count for the active collection.

To inspect actual indexed content, run a representative search, then progressively expand results:

memsearch search "redis ttl"
memsearch expand <chunk_id>

See the CLI reference for stats, and the search / expand sections for content inspection workflows.

Why is search returning irrelevant results?

Common causes:

• the current embedding provider does not match the kind of content you indexed
• the index is stale and needs re-indexing
• your query is too short or too vague
• you switched providers/models and are still searching an old index

Practical things to try:

1. Re-index your memory files: memsearch index . --force
2. Make the query more specific
3. Check which embedding provider you configured
4. If you recently changed providers, consider resetting and rebuilding the index

See Configuration for embedding provider options.

What does "dimension mismatch" mean, and how do I fix it?

A dimension mismatch means your existing Milvus collection was created with one embedding dimension, but your current embedding provider/model is producing vectors with a different dimension.

The usual fix is to reset the collection and re-index from source markdown files:

memsearch reset --yes
memsearch index .

Because markdown is the source of truth, rebuilding the vector index is safe as long as your original memory files are still present.