AuthorsXinyi Fan, Miri Liu, Ruozhen Yang et al.
TL;DR
StateMem uses explicit supersession-aware state units and dependency propagation to cut state drift and raise current-state accuracy to 0.363 on StateMemBench (vs 0.205 Dense retrieval).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
StateMem targets state drift, where 44.4% of failures on LongMemEval oracle are drift even when recall is 1.0 by construction.
These failures show agents with full evidence still answer using superseded values, breaking long-horizon state tracking for tasks like multi-session planning and budgeting.
HOW IT WORKS
StateMem’s core mechanism uses a TurnEncoder to create structured state units in a persistent StateStore, while a deterministic Rechecker tracks supersession and dependencies.
You can think of StateMem like a versioned card catalog: every fact becomes a card, updates create new cards, and dependency edges ensure derived cards get rechecked when inputs change.
This explicit supersession graph lets StateMem maintain the currently operative state under revisions, something a plain context window with raw transcripts cannot reliably do.
DIAGRAM
This diagram shows how StateMem assembles valid state at test time and guides the answer LLM to recompute current values.
DIAGRAM
This diagram shows how StateMemBench generates symbolic programs, renders multi-session dialogues, and scores drift versus gold answers.
PROCESS
Ingestion
StateMem uses the TurnEncoder to parse each conversational turn into state units and writes them into the StateStore with priorities and dependency links.
Update
StateMem applies supersession flags in the StateStore and runs the Rechecker over the dependency graph to mark derived units as needs_recheck.
Test Time
StateMem assembles valid state from active units, including flagged ones, and renders a structured block S grouped by priority and source.
StateMemWrapper
StateMemWrapper traces value chains from retrieved chunks, resolves supersession with precedence rules, and injects a compact state trace before the answer call.
KEY CONTRIBUTIONS
StateMemBench
StateMem introduces StateMemBench with 234 multi-session scenarios and 322 graded probes, generated from symbolic event programs and evaluated with closed-pool drift scoring.
StateMem
StateMem combines the TurnEncoder, StateStore, and Rechecker to explicitly track supersession and dependencies, raising DeepSeek-V4-Flash gold rate from 0.205 to 0.363 on StateMemBench.
StateMemWrapper
StateMemWrapper applies StateMem’s state resolution as a single-call wrapper, adding +31.7 to +66.6 points on StateMemBench over six backends while a matched control attributes +15.0 to +31.7 points to state structure.
RESULTS
State-agreement accuracy
0.363
+0.158 over Dense on DeepSeek-V4-Flash
State-agreement accuracy
0.233
+0.009 over GraphRAG on Qwen-3.5-9B
LongMemEval accuracy
0.656
+0.062 over Mem0 on DeepSeek-V4-Flash
LoCoMo accuracy
0.592
+0.130 over Mem0 on Qwen-3.5-9B
On StateMemBench, which targets state tracking with adversarial traps, StateMem reaches 0.363 gold rate on DeepSeek-V4-Flash and 0.233 on Qwen-3.5-9B. On LongMemEval and LoCoMo, StateMem matches or slightly exceeds long-context baselines while using a bounded state representation, showing that explicit state tracking does not sacrifice recall.
BENCHMARK
On StateMemBench, which targets state tracking with adversarial traps, StateMem reaches 0.363 gold rate on DeepSeek-V4-Flash and 0.233 on Qwen-3.5-9B. On LongMemEval and LoCoMo, StateMem matches or slightly exceeds long-context baselines while using a bounded state representation, showing that explicit state tracking does not sacrifice recall.
BENCHMARK
Gold rate on StateMemBench for Qwen-3.5-9B substrate.
KEY INSIGHT
StateMem shows that long-context models like DeepSeek-V4-Flash score only 0.149 on StateMemBench, while StateMem reaches 0.363 with bounded state.
This is surprising because many assume larger context windows solve memory, but StateMem demonstrates that explicit supersession-aware state tracking matters more than raw context length for evolving state.
WHY IT MATTERS
StateMem unlocks reliable multi-session state tracking by separating recall from supersession-aware state maintenance and making drift a measurable, controllable axis.
Builders can now wrap existing memory systems with StateMemWrapper to gain +31.7 to +66.6 points on state tracking benchmarks, enabling long-running agents that respect current constraints instead of acting on stale decisions.
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.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Zhanghao Hu, Qinglin Zhu et al.
· 2026
xMemory organises agent memory via segments, memory components, groups, and dynamic maintenance with a sparsity plus semantic faithfulness objective. On LoCoMo with GPT-5 nano, xMemory reaches 38.71 BLEU and 50.00 F1 versus Nemori’s 36.65 BLEU and 48.17 F1, while cutting tokens per query from 9154.76 to 6581.20.
Answers use this explainer on Memory Papers.
Checking…