Research Finding

TA-Mem (arXiv:2603.09297, Mar 10, 2026) shows that adaptive tool selection for memory retrieval (choosing exact-key lookup vs. vector similarity per query) outperforms fixed retrieval pipelines for heterogeneous query types.

Applicability

Zeph currently routes all memory retrieval through BM25+RRF hybrid search + MMR re-ranking. This is suboptimal for:

• Episodic queries ("what did I say about X yesterday?") — benefit from exact key/timestamp lookup
• Semantic queries ("find similar concepts to Y") — benefit from vector similarity

Design Sketch

Add a lightweight query classifier before retrieval dispatch in zeph-memory:

1. Classify incoming recall query as episodic (keyword + temporal cues) or semantic
2. Route to appropriate retrieval path:
   • Episodic: SQLite FTS5 + timestamp range filter
   • Semantic: existing BM25+RRF+MMR pipeline
3. No schema changes required — both retrieval paths already exist

The query classifier could be a simple regex/heuristic ("yesterday", "last week", "when did", "remember when") without requiring an LLM call.

Expected Benefit

Improved recall for episodic queries without degrading semantic search quality.

Source

Research session 2026-03-13. arXiv:2603.09297 (TA-Mem).