Skills2026-03-18 · source-backed

EverMemOS — Agent Memory as an Operating System.

Issue

2026-03-18

Confidence

source-backed

Sources

Redaction

passed

Story

Three-phase lifecycle: Phase 1 converts dialogue into MemCells (episodic traces + atomic facts + foresight signals); Phase 2 organizes into MemScenes with conflict resolution; Phase 3 performs MemScene-guided retrieval for minimal sufficient context. State-of-the-art on LoCoMo and LongMemEval benchmarks. arXiv

EverMemOS benchmarked against LongMemEval / Shared entities / Same source domain / Shared topic / What happened next
Unabyss and ZooData top Product Hunt back to back
Linked by a graph relationship (EverMemOS benchmarked against LongMemEval); both cover LoCoMo, LongMemEval; reported by the same outlet (arxiv.org).
EverMemOS benchmarked against LoCoMo / Shared entities / Same source domain / What happened next / Tension
LycheeMemory V2 consolidates at semantic segment level and cuts construction tokens 75.9%.
Linked by a graph relationship (EverMemOS benchmarked against LoCoMo); both cover LoCoMo, LongMemEval; reported by the same outlet (arxiv.org).
EverMemOS benchmarked against LoCoMo / Shared entities / Shared topic / What happened next / Tension
Two agent-memory projects publish directly conflicting benchmark claims on different splits.
Linked by a graph relationship (EverMemOS benchmarked against LoCoMo); both cover LoCoMo, LongMemEval; overlapping topics (benchmark, locomo).
EverMemOS benchmarked against LongMemEval / Shared entities / Shared topic / What happened next
graphify publishes retrieval benchmarks with zero LLM credits: 0.497 recall@10 on LOCOMO, 76% on LongMemEval-S.
Linked by a graph relationship (EverMemOS benchmarked against LongMemEval); both cover LoCoMo, LongMemEval; overlapping topics (benchmark, locomo).
EverMemOS benchmarked against LoCoMo / Shared entity: LoCoMo / Same source domain / What happened next / Tension
Memory system serving cost is unpredictable from conversation length by 18-69%.
Linked by a graph relationship (EverMemOS benchmarked against LoCoMo); both cover LoCoMo; reported by the same outlet (arxiv.org).
EverMemOS benchmarked against LongMemEval / Shared entity: LongMemEval / Shared topic / Earlier coverage
Hindsight: Biomimetic Agent Memory Hits 91% LongMemEval.
Linked by a graph relationship (EverMemOS benchmarked against LongMemEval); both cover LongMemEval; overlapping topics (agent, atomic, fact, longmemeval).
EverMemOS benchmarked against LoCoMo / Shared entities / Shared topic / What happened next
KITE claims SOTA on LoCoMo and LongMemEval with no vectors and ~1.6K tokens of reader context.
Linked by a graph relationship (EverMemOS benchmarked against LoCoMo); both cover LoCoMo, LongMemEval; overlapping topics (agent, benchmark, context, locomo).
EverMemOS benchmarked against LoCoMo / Shared entities / Same source domain / Shared topic / What happened next
AI Agent Memory Costs Explode Without Validated Compaction
Linked by a graph relationship (EverMemOS benchmarked against LoCoMo); both cover LoCoMo, LongMemEval; reported by the same outlet (arxiv.org).

Source trail

arXiv

Entities

Agent Memory EverMemOS LoCoMo LongMemEval MemCells MemScene MemScenes Operating System Phase State

Provenance

Canonical issue: Ramsay Research Agent — March 18, 2026
AI generated: no
Story unit: 2026-03-18-evermemos-agent-memory-as-an-operating-system
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

EverMemOS — Agent Memory as an Operating System.

Story

Related stories

Unabyss and ZooData top Product Hunt back to back

LycheeMemory V2 consolidates at semantic segment level and cuts construction tokens 75.9%.

Two agent-memory projects publish directly conflicting benchmark claims on different splits.

graphify publishes retrieval benchmarks with zero LLM credits: 0.497 recall@10 on LOCOMO, 76% on LongMemEval-S.

Memory system serving cost is unpredictable from conversation length by 18-69%.

Hindsight: Biomimetic Agent Memory Hits 91% LongMemEval.

KITE claims SOTA on LoCoMo and LongMemEval with no vectors and ~1.6K tokens of reader context.

AI Agent Memory Costs Explode Without Validated Compaction

Source trail

Entities

Provenance

EverMemOS — Agent Memory as an Operating System.

Story

Related stories

Unabyss and ZooData top Product Hunt back to back

LycheeMemory V2 consolidates at semantic segment level and cuts construction tokens 75.9%.

Two agent-memory projects publish directly conflicting benchmark claims on different splits.

graphify publishes retrieval benchmarks with zero LLM credits: 0.497 recall@10 on LOCOMO, 76% on LongMemEval-S.

Memory system serving cost is unpredictable from conversation length by 18-69%.

Hindsight: Biomimetic Agent Memory Hits 91% LongMemEval.

KITE claims SOTA on LoCoMo and LongMemEval with no vectors and ~1.6K tokens of reader context.

AI Agent Memory Costs Explode Without Validated Compaction

Source trail

Entities

Provenance