AuthorsZongwei Lv, Yuemeng Xu, Yilun Yao et al.
TL;DR
ArborMem uses a navigable memory forest with interaction-state localization to reach 81.00% on BranchMemEval, +5.00 points over the strongest baseline.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long-term memory benchmarks show that exposing more context does not ensure reliable access, and visibility alone cannot guarantee correct evidence use.
When conversations interleave multiple tasks and plans, assistants misidentify which prior interaction state a new turn resumes, causing thread confusion, lost continuity, and incorrect answers.
HOW IT WORKS
ArborMem combines a Conversation forest, Reusable evidence store, State localization, Context assembly and generation, and Online memory commit to represent resumable interaction trajectories.
You can think of ArborMem like a file system plus an index: each branch is a folder capturing one task’s history, while a separate catalog tracks reusable facts and artifacts.
This design lets ArborMem restore the right branch-local context and add cross-branch evidence, something a plain context window or flat retriever cannot achieve.
DIAGRAM
This diagram shows how ArborMem processes each user turn through state localization, branch-local context reconstruction, cross-branch retrieval, and response generation.
DIAGRAM
This diagram shows how ArborMem is evaluated across benchmarks, read budgets, and ablation configurations.
PROCESS
State Localization
ArborMem uses State localization over the Conversation forest to select a parent state p_t via hybrid retrieval and branch aware reranking.
Context Assembly and Generation
ArborMem reconstructs the branch local trajectory and augments it with evidence from the Reusable evidence store before passing E_t to the Generator LLM.
Online Memory Commit
ArborMem attaches a new node into the Conversation forest and updates the Reusable evidence store with extracted facts and structured records.
Read Budget Constrained Query
ArborMem enforces an answer time evidence budget while selecting branch local context and cross branch evidence for efficient generation.
KEY CONTRIBUTIONS
Interaction-State Localization
ArborMem formalizes State localization over a Conversation forest so each turn first selects which prior trajectory it resumes before retrieving evidence.
Navigable Memory Forest
ArborMem represents long running conversations as a Conversation forest plus Reusable evidence store, separating interaction continuity from reusable facts and artifacts.
BranchMemEval Benchmark
ArborMem introduces BranchMemEval with 33 sessions and 100 questions, showing a 5.00 point gain over the strongest baseline on branch structured memory.
RESULTS
LongMemEval accuracy
68.40%
+9.00 over A-MEM
LoCoMo accuracy
53.15%
+10.31 over BM25
BEAM 100K accuracy
49.75%
+3.36 over Mem0
BranchMemEval accuracy
81.00%
+5.00 over Mem0
On LongMemEval, LoCoMo, BEAM 100K, and BranchMemEval, ArborMem is evaluated for conversational memory and branch structured reasoning. The 81.00% on BranchMemEval proves ArborMem handles interleaved, resumable trajectories better than existing memory systems.
BENCHMARK
On LongMemEval, LoCoMo, BEAM 100K, and BranchMemEval, ArborMem is evaluated for conversational memory and branch structured reasoning. The 81.00% on BranchMemEval proves ArborMem handles interleaved, resumable trajectories better than existing memory systems.
BENCHMARK
Accuracy (%) on BranchMemEval across memory systems.
KEY INSIGHT
On a 50 question LongMemEval subset with a 256 token budget, ArborMem reaches 64.0% accuracy while the next best method reaches only 38.0%.
This is surprising because smaller contexts are usually expected to hurt performance, yet ArborMem’s state localization makes it stronger under the tightest read constraints.
WHY IT MATTERS
ArborMem unlocks reliable handling of interleaved, resumable conversational trajectories by combining interaction state localization with a navigable memory forest.
Builders can now create assistants that track multiple long running tasks and plans over weeks without conflating threads, even under strict context budgets.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…