Source

arXiv:2601.02163 — EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning (January 2026)

Technique

Brain-inspired engram lifecycle in three stages:

1. Trace Formation — converts dialogue turns into MemCells (episodic raw memories)
2. Semantic Consolidation — clusters MemCells into MemScenes with stable user profiles via entity-anchored grouping
3. Reconstructive Recollection — scene-guided retrieval assembles minimal sufficient context from MemScenes rather than raw messages

Achieves SOTA on LoCoMo and LongMemEval benchmarks. Key property: consolidation runs offline (async), recollection is a single scene→context reconstruction step (fast path).

Applicability to Zeph

High. Zeph's session digest (PR #2309) compresses recent turns but doesn't consolidate into stable semantic structures. The MemScene layer maps directly to Zeph's memory tier system: MemCells ≈ raw SQLite messages, MemScenes ≈ promoted semantic clusters. The scene-guided recollection would replace or augment the current MMR retrieval pass.

Differs from:

• research(memory): A-MAC adaptive memory admission control — 5-factor value decomposition for principled fact filtering (arXiv:2603.04549) #2317 (A-MAC) — admission filtering on what to write, not consolidation structure
• research(memory): All-Mem lifelong memory — online fast-path writes + offline topology-edit consolidation, no destructive summarization (arXiv:2603.19595) #2270 (All-Mem) — write-path topology (online fast + offline consolidation is similar, but All-Mem focuses on avoiding destructive summarization; EverMemOS focuses on scene construction)

Implementation sketch

1. Add MemScene struct to zeph-memory — a group of related MemCells with a stable entity profile
2. Background consolidator (async fn consolidate_scenes) runs after tier promotion sweep, merges semantically related MemCells into MemScene records
3. Recollection: when assembling context, retrieve relevant MemScenes first, then expand into MemCells as needed

P2 — research, evaluate fit with existing tier + graph memory architecture before implementation.