Description

MemMachine presents a ground-truth-preserving personalized memory system that stores full conversational episodes alongside processed summaries. Key finding: retrieval-stage optimization substantially outperforms ingestion-stage optimization.

Key Insights

Retrieval improvements (total ~9.5%):

• Retrieval depth tuning: +4.2%
• Context formatting: +2.0%
• Search prompt design: +1.8%
• Query bias correction: +1.4%

Ingestion improvements:

• Sentence chunking: +0.8% (much smaller gain)

Conclusion: how you retrieve stored information matters far more than how you chunk/process it at ingestion time.

Efficiency: 80% fewer input tokens vs. competing systems; 93% accuracy on LongMemEvalS.

Three Memory Layers

1. Short-term (in-context)
2. Long-term episodic (full conversation episodes preserved)
3. Profile memory (distilled user model)

Relevance to Zeph

Zeph's uses summarization at write time (ingestion-side). MemMachine's findings suggest that investing more in retrieval-side tuning (search prompt design, depth, formatting, query bias correction) would yield higher returns than improving the summarization pipeline.

Concrete actions:

• Add retrieval depth config: number of candidates to fetch before MMR/reranking
• Improve search prompt templates sent to the embedding/retrieval model
• Add query bias correction: detect when query is first-person vs. topic query and adjust embedding
• Preserve raw episode data alongside summaries (current impl may summarize-only)

Source

• Paper: https://arxiv.org/abs/2604.04853
• Benchmark: LongMemEvalS