AuthorsRunzhe Wang, Huilin Lu, Shengjie Liu et al.
TL;DR
AdMem uses a bi-level semantic episodic procedural memory with actor critic memory agents and achieves 63.4% vs 49.3% task completeness on AlfWorld over ReAct (+14.1pp).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
AdMem targets long-horizon tasks where agents must remember, organize, and reuse knowledge, but prior memory mainly stores factual information and replays past successes.
In environments like AgentBoard, this causes brittle behavior on multi-turn tool-use tasks, where agents repeatedly fail at critical steps without learning from those failures.
HOW IT WORKS
AdMem’s core mechanism combines Actor agent, Long-term memory agent, Critic agent, Short-term memory, and Reward-based long-term memory management into a single POMDP-style agent.
You can think of AdMem like RAM plus disk plus a coach: Short-term memory is the stack-like RAM, long-term memory is disk, and the Critic agent is a coach labeling good and bad habits.
This design lets AdMem learn which procedural memories actually help, prune or merge unhelpful ones, and adaptively retrieve guidance that a plain context window or static replay cannot provide.
DIAGRAM
This diagram shows how AdMem’s Actor agent, Critic agent, and Long-term memory agent interact each turn to generate actions and update memories.
DIAGRAM
This diagram shows how AdMem is evaluated on AgentBoard domains and decomposed into ablations with and without short-term and long-term procedural memory.
PROCESS
Set up
In the Set up phase, AdMem formalizes the task-solving environment as a POMDP and initializes the Actor agent state and Short-term memory.
Agent Pipeline
During Agent Pipeline execution, the Actor agent queries the Long-term memory agent and sends half-baked memories to the Critic agent while acting.
Short term memory for context compaction
In Short term memory for context compaction, AdMem uses a stack-like Short-term memory with sub-goals to summarize irrelevant history and keep context focused.
Reward-based long-term memory management
In Reward-based long-term memory management, the Long-term memory agent updates procedural memory scores with the EM-based reward model and prunes or merges entries.
KEY CONTRIBUTIONS
A comprehensive memory system
AdMem introduces a comprehensive memory system that supports generation, storage, management, and retrieval of semantic, episodic, and procedural memories via the Long-term memory agent and Short-term memory.
Agent planning paradigm for memory generation
AdMem designs an agent planning paradigm with tools like Plan for the new task and Add sub-goal so the Actor agent and Critic agent can generate rich procedural memories.
Reward-based long-term memory management
AdMem proposes Reward-based long-term memory management that uses a bandit-style EM model over vm parameters to score procedural entries and adapt retrieval and pruning.
RESULTS
Task completeness AlfWorld
63.4%
+14.1 percentage points over ReAct
Average progress AlfWorld
0.7755
+0.1169 over ReAct
Task completeness Babyai
100.0%
+2.7 percentage points over ReAct
Task completeness Jericho third epoch
60.0%
+20.0 percentage points over LLM + LTP
On AgentBoard domains such as AlfWorld, Babyai, and Jericho, AdMem is evaluated for task completeness and average progress, showing that reward-driven procedural memory improves long-horizon task-solving over ReAct and AWM baselines.
BENCHMARK
On AgentBoard domains such as AlfWorld, Babyai, and Jericho, AdMem is evaluated for task completeness and average progress, showing that reward-driven procedural memory improves long-horizon task-solving over ReAct and AWM baselines.
BENCHMARK
Task completeness C on AlfWorld from AgentBoard.
BENCHMARK
Task completeness C on Jericho over three epochs for procedural memory ablations.
KEY INSIGHT
Adding naive long-term procedural memory alone (LLM + LTP) initially reduces Jericho task completeness from 40.0% to 20.0% in the first epoch.
This is surprising because more experience should help, but without AdMem’s reward-based management, extra trajectories create misleading guidance and hurt performance before slowly recovering.
WHY IT MATTERS
AdMem shows that agents can maintain scalable semantic, episodic, and procedural memories that are continuously evaluated and improved online with rewards.
Builders can now deploy long-lived agents that not only store past workflows but also learn which procedures actually help, pruning bad habits and reusing effective strategies across tasks.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…