What happened?

tool_check_duplicate in mcp_server.py and Layer3.search() in layers.py can return negative similarity scores for very dissimilar content.

Both use round(1 - dist, 3) to convert a ChromaDB cosine distance into a similarity score. With hnsw:space=cosine, ChromaDB distances are in the range [0, 2] — not [0, 1]. For maximally dissimilar vectors the distance slightly exceeds 1.0, making 1 - dist negative.

The rest of the codebase already has the correct pattern: searcher.py line 285 uses round(max(0.0, 1 - dist), 3). The two affected sites are missing the max(0.0, ...) clamp.

Affected lines:

• mempalace/mcp_server.py — tool_check_duplicate(): similarity = round(1 - dist, 3)
• mempalace/layers.py — Layer3.search(): similarity = round(1 - dist, 3)

What did you expect?

Similarity scores should always be in [0.0, 1.0]. A score of -0.004 is meaningless and could confuse AI clients that read or display the value (e.g. tool_check_duplicate returns the similarity in its JSON response, and Layer3 renders it as (sim=-0.004) in the memory context block).

How to reproduce:

1. Install mempalace and run the snippet below (no palace required — uses an in-memory ChromaDB client).

2. Run:

import chromadb

client = chromadb.Client()
col = client.get_or_create_collection("bug_demo", metadata={"hnsw:space": "cosine"})

col.add(
    ids=["drawer_1"],
    documents=["The Pythagorean theorem states that a^2 + b^2 = c^2 in a right triangle."],
    metadatas=[{"wing": "math", "room": "geometry"}],
)

results = col.query(
    query_texts=["Chocolate cake recipe with vanilla frosting and strawberries."],
    n_results=1,
    include=["distances"],
)

dist = results["distances"][0][0]
print(f"distance  : {dist:.4f}")           # e.g. 1.0035
print(f"similarity: {round(1 - dist, 3)}") # e.g. -0.004  ← negative!

3. Observe that similarity is negative.

Output:

distance  : 1.0035
similarity: -0.004

Environment:

• OS: macOS
• Python version: 3.11.9
• MemPal version: git SHA 2792ce8 (develop)