AgeMem: Unified Long- and Short-Term Memory for LLM Agents (2026)

Free

First to unify LTM (add/update/delete) and STM (retrieve/summarize/filter) as tool-based actions via GRPO RL; 7B model achieves +49.59% over no-memory baseline across 5 benchmarks; ICLR 2026 MemAgents Workshop

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About AgeMem: Unified Long- and Short-Term Memory for LLM Agents (2026)

Agentic Memory (AgeMem) is a unified framework for managing long-term memory (LTM) and short-term memory (STM) in large language model (LLM) agents. It exposes memory operations—such as storing, retrieving, updating, summarizing, and discarding information—as tool-based actions that the agent can autonomously decide when to invoke. To train these unified behaviors, AgeMem employs a three-stage progressive reinforcement learning strategy with step-wise GRPO to handle sparse and discontinuous rewards. Evaluated on five long-horizon benchmarks, AgeMem consistently outperforms strong memory-augmented baselines, achieving improved task performance, higher-quality long-term memory, and more efficient context usage. The approach is open-source with code available, and a 7B parameter model version demonstrated a 49.59% improvement over a no-memory baseline. AgeMem was accepted as an ACL 2026 SAC Highlight.

Key Features

Unified long-term and short-term memory management as tool-based actions

Autonomous decisions on when to store, retrieve, update, summarize, or discard information

Three-stage progressive reinforcement learning with step-wise GRPO

Open-source code available

Improves task performance, memory quality, and context efficiency

Evaluated on five long-horizon benchmarks

Pros & Cons

Pros

First unified approach for LTM and STM management within the LLM agent's policy
Significant performance gains over no-memory and memory-augmented baselines (49.59% improvement)
Open-source with publicly available code
Efficient context usage reducing strain on finite context windows

Cons

Requires reinforcement learning training which can be computationally expensive
Primarily validated on 7B parameter models; scalability to larger models is untested
Implementation complexity due to multi-stage training pipeline
As a research framework, not yet packaged as a plug-and-play tool

Best For

Long-horizon reasoning tasks for LLM agentsApplications requiring adaptive memory management across extended interactionsScenarios where finite context windows limit agent performanceResearch on reinforcement learning for memory-augmented LLM agents

FAQ

What is AgeMem?

AgeMem (Agentic Memory) is a unified framework that integrates long-term and short-term memory management directly into the policy of LLM agents, using tool-based actions learned via GRPO reinforcement learning.

How does AgeMem work?

AgeMem exposes memory operations (store, retrieve, update, summarize, discard) as tools that the LLM agent can autonomously invoke. It employs a three-stage progressive reinforcement learning strategy with step-wise GRPO to train these behaviors.

Is the code available?

Yes, the code for AgeMem is publicly available and linked from the arXiv paper.

What benchmarks were used for evaluation?

AgeMem was evaluated on five long-horizon benchmarks, consistently outperforming strong memory-augmented baselines.

What was the performance improvement?

A 7B parameter model using AgeMem achieved a 49.59% improvement over a no-memory baseline across multiple benchmarks.