Title: Search does not work for non-English languages (Russian, etc.) — default embedding model is English-only

[Speccy-Rom]

Problem
▎
▎ MemPalace uses ChromaDB with the default embedding function (all-MiniLM-L6-v2 from sentence-transformers). This model is trained primarily on English text and produces near-random vectors for non-English
▎ languages (Russian, Chinese, Arabic, etc.).
▎
▎ As a result, semantic search queries in Russian return irrelevant results with very low similarity scores (0.04–0.18), making the tool practically unusable for non-English projects or users who store
▎ memories in their native language.
▎
▎ Steps to reproduce
▎ 1. Mine a project that contains Russian-language text, or store a drawer with Russian content via mempalace_add_drawer
▎ 2. Search for it in Russian: mempalace search "запрос на русском"
▎ 3. Results are unrelated, similarity scores < 0.2
▎
▎ Expected behavior
▎
▎ Search should work across languages, or at minimum there should be a way to configure a multilingual embedding model.
▎
▎ Suggested fix
▎
▎ Allow configuring the embedding model in ~/.mempalace/config.json, defaulting to a multilingual model such as paraphrase-multilingual-mpnet-base-v2 (supports 50+ languages, similar size to all-MiniLM-L6-v2).
▎ Alternatively, use multilingual-e5-small for a lighter option.
▎
▎ {
▎ "embedding_model": "paraphrase-multilingual-mpnet-base-v2"
▎ }

[marcuss-grindr]

this could be auto-resolve with my latest request for enhancement #756

[codysnider]

Try https://github.com/codysnider/tagmem

It supports multi and cross-lingual storage and search: https://github.com/codysnider/tagmem/blob/main/MULTILINGUAL_RETRIEVAL.md

It also outperforms mempalace in all benchmarks.

[NickShtefan]

This is exactly what PR #442 solves.

I've been running it in production on my own palace with Russian content for several weeks now. The difference is dramatic:

• Before (default all-MiniLM-L6-v2): search scores 0.19–0.40 for Russian queries — essentially random
• After (intfloat/multilingual-e5-base): scores 0.70–0.77 — actually useful results

The embedding model is bound to the palace at init time via ChromaDB collection metadata — no env vars needed after initial setup. Includes mismatch detection so you can't accidentally query with the wrong model, plus a re-mine command for safe migration.

10K+ drawers mined, daily use with MCP server + Claude Code. Waiting on review from @igorls — addressed his feedback (removed env var config, moved to init-time binding). Rebased on develop as of today.

@Speccy-Rom — if you want to try it before merge: pip install -e ".[multilingual]" from my fork, then mempalace re-mine with the model configured.