Current Behavior
optimize_query is currently only applied when search_mode == "semantic". For keyword and hybrid modes, it is silently ignored.
The gating logic is in long_term_memory.py:
# Optimize query only for semantic search; keyword and hybrid modes should
# preserve the literal text for lexical matching.
if optimize_query and text and search_mode == SearchModeEnum.SEMANTIC:
search_query = await optimize_query_for_vector_search(text)The current optimization uses an LLM to rewrite conversational queries into embedding-friendly form:
"tell me what I like to eat" → "user food preferences dietary likes dislikes favorite meals cuisine"
This is effective for vector similarity but would hurt BM25/keyword matching since the rewritten terms may not appear in stored text.
Why This Matters
- Users setting
optimize_query=True with search_mode="keyword" or "hybrid" get no optimization with no warning
- Hybrid mode uses both vector and keyword components — the vector half could benefit from optimization while the keyword half uses the original text
Options to Consider
-
Do nothing — current behavior is intentional and documented (docstrings now clarify this is semantic-only). Keyword search needs the original terms for lexical matching.
-
Hybrid-aware optimization — for hybrid mode, optimize the query for the vector component but pass the original text to the keyword component. This would require search_memories to accept two query strings (one optimized for embedding, one original for BM25).
-
Keyword-specific optimization — create a separate optimization path for keyword search (e.g., expanding synonyms, stripping filler words while keeping searchable terms) instead of the current vector-focused rewrite.
-
Warn when ignored — log a warning or raise when optimize_query=True is passed with a non-semantic mode, so users know it's not being applied.
Related Files
agent_memory_server/long_term_memory.py — gating logic (line 1187)agent_memory_server/llm/client.py — optimize_query_for_vector_search()agent_memory_server/config.py — query_optimization_prompt_template
Current Behavior
optimize_queryis currently only applied whensearch_mode == "semantic". For keyword and hybrid modes, it is silently ignored.The gating logic is in
long_term_memory.py:The current optimization uses an LLM to rewrite conversational queries into embedding-friendly form:
"tell me what I like to eat"→"user food preferences dietary likes dislikes favorite meals cuisine"This is effective for vector similarity but would hurt BM25/keyword matching since the rewritten terms may not appear in stored text.
Why This Matters
optimize_query=Truewithsearch_mode="keyword"or"hybrid"get no optimization with no warningOptions to Consider
Do nothing — current behavior is intentional and documented (docstrings now clarify this is semantic-only). Keyword search needs the original terms for lexical matching.
Hybrid-aware optimization — for hybrid mode, optimize the query for the vector component but pass the original text to the keyword component. This would require
search_memoriesto accept two query strings (one optimized for embedding, one original for BM25).Keyword-specific optimization — create a separate optimization path for keyword search (e.g., expanding synonyms, stripping filler words while keeping searchable terms) instead of the current vector-focused rewrite.
Warn when ignored — log a warning or raise when
optimize_query=Trueis passed with a non-semantic mode, so users know it's not being applied.Related Files
agent_memory_server/long_term_memory.py— gating logic (line 1187)agent_memory_server/llm/client.py—optimize_query_for_vector_search()agent_memory_server/config.py—query_optimization_prompt_template