mcp-memory-service: exact text only scores 60-70%? Use hybrid BM25 + vector search (v10.8.0)

Question

In mcp-memory-service (doobidoo), searching with text that exactly matches a stored document chunk returns similarity scores around 60-70% instead of the expected 90-100%. Vector embeddings are affected by normalization, context window, and floating-point precision, so identical text still gets a low score. Is there a way to get near-perfect scores for exact text matches? The fix landed in v10.8.0 — how do I enable and configure the hybrid BM25 + vector search?

Originally asked on GitHub: https://github.com/doobidoo/mcp-memory-service/issues/175

AgentArmorHQ team · Accepted Answer

**Status: fixed in v10.8.0** (released 2026-02-08). The maintainer implemented hybrid BM25 + vector search: FTS5-based BM25 keyword search with a trigram tokenizer runs in parallel with the vector search and the scores are fused. Exact matches that previously scored 60-70% now score near-100%.

Upgrade and switch the search mode:

```bash
pip install --upgrade mcp-memory-service
```

```bash
curl -X POST http://127.0.0.1:8000/api/search \
  -H "Content-Type: application/json" \
  -d '{"query": "OAuth 2.1 authentication", "mode": "hybrid", "limit": 10}'
```

Python API:

```python
results = await storage.search_memories(query="exact text to find", mode="hybrid", limit=5)
# or directly:
results = await storage.retrieve_hybrid(query="specific keywords", n_results=10)
```

Configuration (hybrid is enabled by default; weights should sum to ~1.0):

```bash
export MCP_HYBRID_SEARCH_ENABLED=true
export MCP_HYBRID_KEYWORD_WEIGHT=0.3    # BM25 keyword weight
export MCP_HYBRID_SEMANTIC_WEIGHT=0.7   # vector similarity weight
```

The maintainer lists 0.5/0.5 as more keyword-focused, 0.2/0.8 as more semantic-focused, and 0.3/0.7 as the recommended default. `mode="semantic"` is unchanged, so the release is backward compatible. Reported latency is <50ms for 100 memories.

A commenter who built the same pattern in another project (AgentKits Memory) confirmed SQLite's FTS5 BM25 is sufficient and stressed normalizing scores before merging: FTS5 `bm25()` returns negative values where closer to 0 is better, so they used `min(1, max(0, 1 + bm25_rank / 10))`, and `1 - cosine_distance` for the vector side; otherwise one score dominates. They also made the weight adaptive by query type — exact strings like error codes or function names get higher text weight, while natural-language questions get higher vector weight. General advice: if you implement your own hybrid search, normalize both scores to 0-1 before combining.

AgentArmorHQ team · Answer

**What a GitHub user tried on GitHub** (Outcome unclear)

The issue notes that v8.6.1 already switched from L2 distance to cosine similarity as an alternative; the issue describes it as improving scores but not solving the exact-match problem, without test results.

Source: https://github.com/doobidoo/mcp-memory-service/issues/175

AgentArmorHQ team · Answer

**What @leduclinh7141 tried on GitHub** (Worked)

A commenter implemented hybrid search in their own project using SQLite FTS5 (BM25) and HNSW vector search in parallel, normalizing both scores to 0-1 before merging; they reported exact matches jumped to near-100% in the merged results.

Source: https://github.com/doobidoo/mcp-memory-service/issues/175#issuecomment-3846723977

AgentArmorHQ team · Answer

**What @doobidoo tried on GitHub** (Worked)

The maintainer announced hybrid search was implemented and released in v10.8.0: FTS5 BM25 with trigram tokenizer, parallel execution, and configurable fusion weights (default 30% keyword / 70% semantic).

Source: https://github.com/doobidoo/mcp-memory-service/issues/175#issuecomment-3866270459

mcp-memory-service: exact text only scores 60-70%? Use hybrid BM25 + vector search (v10.8.0)

4 answers

Add your answer or comment

mcp-memory-service: exact text only scores 60-70%? Use hybrid BM25 + vector search (v10.8.0)

4 answers

Add your answer or comment

Related questions