AuthorsTianyu Yang, Sudipta Paul, Vijay Srinivasan et al.
TL;DR
TRUSTMEM uses a Memory Transition Verifier plus Transition-Ranked GRPO to train memory updates, reaching 65.7 average on MemoryAgentBench (+6.5 over Mem-α).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM memory agents can omit key facts, corrupt prior knowledge, or hallucinate unsupported content during consolidation, which then persists across interactions.
These persistent errors contaminate long-term memory, degrading downstream reasoning and question answering even when terminal task performance looks acceptable.
HOW IT WORKS
TRUSTMEM centers on a Memory Transition Verifier, Memory Executor, Policy Model, Reference Model, and Retriever to score and update memory transitions step by step.
Think of TRUSTMEM like a CPU with a safety coprocessor: the Policy Model proposes edits, the Memory Executor applies them, and the Memory Transition Verifier checks each change before it becomes disk.
This verifier plus Transition-Ranked GRPO lets TRUSTMEM learn memory updates that avoid omission, corruption, and hallucination, something a plain context window or trajectory-only reward cannot guarantee.
DIAGRAM
This diagram shows how TRUSTMEM processes each input chunk, executes WRITE REVISE PRUNE actions, and verifies the resulting memory transition before downstream use.
DIAGRAM
This diagram shows how TRUSTMEM uses multi-granularity rewards and verifier-guided ranking to optimize the memory-update policy.
PROCESS
Continual memory update problem
TRUSTMEM receives chunk ct and previous memory Mt−1, then the Policy Model proposes structured WRITE REVISE PRUNE actions for this step.
Memory Executor applies actions
The Memory Executor deterministically applies actions to the Memory Store, producing an updated memory state Mt and defining the transition zt.
Memory Transition Verifier scoring
The Memory Transition Verifier inspects zt for coverage, preservation, and faithfulness, returning a trustworthiness score used as Rver in training.
Transition-Ranked GRPO optimization
TRUSTMEM runs multiple rollouts, uses verifier scores to rank transitions, and updates the Policy Model with GRPO plus a DPO-style ranking loss.
KEY CONTRIBUTIONS
Memory Transition Verifier
TRUSTMEM introduces a Memory Transition Verifier that scores each update for coverage, preservation, and faithfulness, cutting omission by 40.1% and corruption by 79.1% versus baselines.
Transition-Ranked GRPO
TRUSTMEM proposes Transition-Ranked GRPO, using verifier-guided preference pairs to optimize memory-editing actions beyond scalar rewards, boosting TTL scores from 57.6 to 70.1.
Reliable long-term memory benchmarks
TRUSTMEM is evaluated on MemoryAgentBench, HaluMem, and Mem-α, achieving 65.7 average on MemoryAgentBench and +12.14 F1 on HaluMem memory extraction over the strongest prior method.
RESULTS
MemoryAgentBench Avg.
65.7 score
+6.5 over Mem-α
Mem-α Avg.
66.3 score
+4.4 over Mem-α baseline
HaluMem Extraction F1
69.45 F1
+12.14 over strongest prior F1
Omission Rate
8.08 %
lowest omission vs MemAgent, MEM1, Mem-T, AtomMem
On MemoryAgentBench, which tests accurate retrieval, test-time learning, and long-range understanding, TRUSTMEM’s 65.7 average score shows robust memory utility. On HaluMem, TRUSTMEM’s 69.45 F1 for memory extraction and 8.08% omission rate demonstrate safer, higher-quality memory operations than prior memory agents.
BENCHMARK
On MemoryAgentBench, which tests accurate retrieval, test-time learning, and long-range understanding, TRUSTMEM’s 65.7 average score shows robust memory utility. On HaluMem, TRUSTMEM’s 69.45 F1 for memory extraction and 8.08% omission rate demonstrate safer, higher-quality memory operations than prior memory agents.
BENCHMARK
Average task score across Accurate Retrieval, Test-Time Learning, and Long-Range Understanding on MemoryAgentBench.
BENCHMARK
Average score on the Mem-α validation set across SQuAD, HotpotQA, PerLTQA, TREC-C, NLU, PubMed, and BookSum.
KEY INSIGHT
TRUSTMEM’s Transition-Ranked GRPO boosts Test-Time Learning from 57.6 to 70.1, even though both variants share the same reward components.
This is surprising because many assume scalar rewards alone suffice; TRUSTMEM shows that explicit transition ranking adds crucial supervision for safe memory updates.
WHY IT MATTERS
TRUSTMEM unlocks trustworthy, step-level memory consolidation, letting agents audit each WRITE REVISE PRUNE transition instead of only checking final answers.
Builders can now deploy long-lived LLM agents whose memories evolve safely over thousands of interactions, without accumulating silent omissions, corruptions, or hallucinations.
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.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…