SeoPalace: an independent lexical layer on top of MemPalace (and a question about the name)

[69360albin-oss]

Hi all,

I use MemPalace every day as the semantic memory behind my personal knowledge base (about 1,300 Markdown notes, in French). I built a small open-source layer on top of it and I would rather you hear about it from me than from someone else, especially because of the name.

Repo: https://github.com/69360albin-oss/SeoPalace (MIT)

What it is

• One "## Keywords" line at the bottom of each note, written the way I actually type when I search: terms, synonyms, and full phrases.
• A dozen search phrases generated per note by an LLM (doc2query), kept only if MemPalace alone, queried with that phrase, brings back the original note. On my base 39.5% of the generated phrases were thrown away by that filter.
• A small fielded BM25 index over title / keywords / phrases / body, fused with MemPalace results through RRF.
• The MemPalace adapter only calls your public search_memories API. MemPalace is an optional backend; nothing of yours is vendored, modified or rebranded.

What I measured (one user, one private base, so take it for what it is)
On 94 questions taken from real usage (queries my agents actually sent, and my own messages):

• MemPalace alone: right note first 18/94, in the top 3 31/94
• Keywords line + MemPalace (RRF): 32/94 and 52/94
• plus filtered generated phrases: 33/94 and 55/94

I know MemPalace already does hybrid retrieval internally, so this is not "MemPalace is missing BM25". My reading is that the gain comes from the note side: the notes are written in the vocabulary of the solution, and I search them in the vocabulary of the symptom. The README also shows a public benchmark (200 fictional notes) where MemPalace alone beats my fusion, 48/55 vs 42/55, and a list of everything that did not work (CPU cross-encoder reranker, MaxP, graph authority, and so on). Both tables are there on purpose.

Two questions for you

1. The name. "SeoPalace" is obviously a nod to MemPalace, and I have read your warning about impostor sites. The README opens with a notice that the project is independent and not affiliated with you, links to your official repo, and uses none of your branding. If the name is still a problem for you, tell me and I will rename it. No hard feelings.
2. If the fusion or the coherence filter is done in a way that misuses your API, or if there is a better entry point than search_memories for this, I would like to know.

Thanks for MemPalace. It is the part of my setup that works without me thinking about it.