Paper

arXiv:2603.14517 — Learning to Forget: Sleep-Inspired Memory Consolidation for Resolving Proactive Interference in LLMs (March 2026)

Key Finding

SleepGate augments memory with a learned sleep cycle consisting of three phases:

1. Synaptic downscaling: reduces activation strength of all memories proportionally
2. Selective replay: replays high-utility, high-recency traces to restore importance
3. Targeted forgetting: prunes memories that fall below a utility threshold after downscaling

Result: retrieval interference horizon drops from O(n) to O(log n). At PI (proactive interference) depth where all baselines collapse below 18% retrieval accuracy, SleepGate achieves 97–99.5%.

Applicability to Zeph

Directly addresses a known failure mode: long-running Zeph sessions accumulate stale entity associations in MAGMA that corrupt recent-fact retrieval. For example, if the user mentions "Alice works at Google" in session 10, then "Alice moved to Microsoft" in session 50, the old edge can still surface with higher activation than the new one.

The consolidation sweep in zeph-memory (All-Mem, MemScene) already runs periodically. SleepGate's insight is to add an explicit forgetting phase to the consolidation loop:

• After the similarity merge pass (deduplicate near-identical memories)
• Apply a utility decay to all items not recently activated
• Prune items below a minimum utility floor

Integration Point

crates/zeph-memory/src/consolidation.rs — add a forgetting_sweep() step after consolidation_sweep(). Config: [memory.consolidation] forgetting_enabled = true, forgetting_decay_rate = 0.1, forgetting_floor = 0.05.

Related

• research(memory): comprehensive memory survey — write-manage-read taxonomy, 5 mechanism families, open challenges (arXiv:2603.07670) #2351 (memory survey) — "learned forgetting" is explicitly listed as an open challenge
• feat(memory): All-Mem consolidation, MAGMA edge weights, RuntimeLayer hooks #2358 (All-Mem consolidation) — the sweep infrastructure is already in place