Public story · 2026-02-28 · source-backed

EMPO2: Memory-Augmented Agent (ICLR 2026)

Issue

2026-02-28

Confidence

source-backed

Sources

Redaction

redacted

Story

Hybrid on/off-policy RL giving agents non-parametric memory for exploration. 128.6% improvement over GRPO on ScienceWorld, 11.3% on WebShop. Agents generalize to out-of-distribution tasks with "only a few trials with memory and no parameter updates." (arXiv)

SkillPyramid benchmarked against ScienceWorld / Shared entities / Same source domain / Shared topic / What happened next
SkillPyramid organizes agent skills into a hierarchy
Linked by a graph relationship (SkillPyramid benchmarked against ScienceWorld); both cover ScienceWorld, WebShop; reported by the same outlet (arxiv.org).
GRPO competes with PPO / Shared entities / Same source domain / Shared topic / What happened next
Memory-R1: RL-Based Agent Memory Management
Linked by a graph relationship (GRPO competes with PPO); both cover GRPO, Memory; reported by the same outlet (arxiv.org).
Unsloth released GRPO / Shared entity: GRPO / What happened next / Tension
Unsloth makes single-H100 long-context RL fine-tuning actually tractable.
Linked by a graph relationship (Unsloth released GRPO); both cover GRPO; picks up the GRPO thread on 2026-06-20.
Shared entities / Same source domain / Shared topic / What happened next
HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.
Both cover GRPO, WebShop; reported by the same outlet (arxiv.org); overlapping topics (agent, grpo, over).
DeepSeek-R1 uses GRPO
Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE
Linked by a graph relationship (DeepSeek-R1 uses GRPO).
Microsoft released EMPO2 / Shared entity: Memory / Shared topic / What happened next
Agent Infrastructure Fills Five Of Product Hunt's Top Seven
Linked by a graph relationship (Microsoft released EMPO2); both cover Memory; overlapping topics (agent, memory).
EMPO2 benchmarked against ScienceWorld / Shared entity: ScienceWorld / Same source domain / What happened next
SkillReranker improves accuracy and cuts cost at the same time.
Linked by a graph relationship (EMPO2 benchmarked against ScienceWorld); both cover ScienceWorld; reported by the same outlet (arxiv.org).
Shared entity: Memory / Same source domain / Shared topic / What happened next / Tension
110ms Memory Delay Makes AI Agents Repeat Actions
Both cover Memory; reported by the same outlet (arxiv.org); overlapping topics (agent, memory).

Source trail

arXiv

Entities

Augmented Agent EMPO2 GRPO Hybrid ICLR Memory ScienceWorld WebShop

Provenance

Canonical issue: Ramsay Research Agent — 2026-02-28
AI generated: no
Story unit: 2026-02-28-empo2-memory-augmented-agent-iclr-2026
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

EMPO2: Memory-Augmented Agent (ICLR 2026)

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Agent Infrastructure Fills Five Of Product Hunt's Top Seven

SkillReranker improves accuracy and cuts cost at the same time.

110ms Memory Delay Makes AI Agents Repeat Actions

Source trail

Entities

Provenance

EMPO2: Memory-Augmented Agent (ICLR 2026)

Story

Related stories

SkillPyramid organizes agent skills into a hierarchy

Memory-R1: RL-Based Agent Memory Management

Unsloth makes single-H100 long-context RL fine-tuning actually tractable.

HCAPO: Hindsight Credit Assignment for Sparse-Reward Agent RL.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Agent Infrastructure Fills Five Of Product Hunt's Top Seven

SkillReranker improves accuracy and cuts cost at the same time.

110ms Memory Delay Makes AI Agents Repeat Actions

Source trail

Entities

Provenance