Finding

Paper: "Beyond the Context Window: A Cost-Performance Analysis of Fact-Based Memory vs. Long-Context LLMs for Persistent Agents"
arXiv: https://arxiv.org/abs/2603.04814 (March 2026)

Core Idea

Empirically compares structured fact-extraction memory (Mem0-style SQLite/vector store) against full-history long-context inference on three persistence benchmarks (LongMemEval, LoCoMo, PersonaMemv2):

• Long-context wins on recall accuracy at short histories
• Fact-based memory wins dramatically on cost at scale
• Provides a crossover formula: at what message count does each approach dominate

Applicability to Zeph (5/5)

Zeph already implements exactly this tradeoff (SQLite episodic + Qdrant semantic vs. full context). The paper's crossover formula directly justifies Zeph's memory-first design and could inform:

1. A configurable field in — uses the crossover formula to pick at runtime
2. Cost tracking already in Zeph can be extended with a running cost-per-turn estimate to trigger memory-first mode when cost threshold is hit
3. Validates the compaction + semantic retrieval pipeline as cost-optimal for long sessions

Implementation Sketch

In : add a strategy selector that queries current turn count + cost metrics, switches from full-history to memory-retrieval-first assembly when the crossover threshold is exceeded.