Paper

Adaptive Memory Admission Control for LLM Agents (A-MAC)

• arXiv: https://arxiv.org/abs/2603.04549 (submitted 4 March 2026, ICLR 2026 Workshop MemAgent)

Summary

A-MAC treats memory admission as a structured decision problem, decomposing "memory value" into five interpretable factors: future utility, factual confidence, semantic novelty, temporal recency, and content type prior. Rather than opaque LLM-driven policies that accumulate hallucinated or stale facts, A-MAC provides lightweight, auditable admission gates on what enters long-term storage.

Relevance to Zeph

zeph-memory currently writes facts to SQLite/Qdrant without principled admission control — any assistant response can be saved via memory_save. This framework maps directly to filtering what gets persisted after each session:

• future_utility ↔ existing importance_weight config in [memory.semantic] (partial overlap, but A-MAC is more principled)
• factual_confidence — not implemented; would require a lightweight verifier on write
• semantic_novelty — partially covered by MMR at recall time, but not at write time
• temporal_recency — covered by temporal_decay_half_life_days
• content_type_prior — not implemented; different handling for code vs. facts vs. opinions

A write-time admission gate would reduce memory bloat, stale-fact accumulation, and hallucination propagation into future sessions.

Priority

P2 — addresses a known gap (unbounded memory growth, stale facts) with a concrete multi-factor framework.