AuthorsZheling Tan, Jin Gao, Dequan Wang
TL;DR
CABLE uses complementary antecedent-based linking plus bounded expansion to extend semantic retrieval, boosting A-MEM from 59.33% to 65.33% on MA-LongMemEval (+6.00 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CABLE targets cases where stored evidence is present but unreachable because retrieval relies on semantic similarity and misses semantically distant antecedents.
In long-term conversational memory, A-MEM and similar systems can fail when later questions require earlier experiences or motivations, leaving agents unable to answer despite having the needed history.
HOW IT WORKS
CABLE combines Antecedent-Oriented Query Generation, Dual Retrieval, Overlap Subtraction, and Verification and Graph Update to build sparse, retriever-complementary links between memories.
Think of CABLE as adding a curated cross-reference catalog on top of a search index, so agents can jump to causes and motivations that plain similarity search overlooks.
By storing and reusing these antecedent links, CABLE lets long-term memory systems recover semantically distant but reasoning-critical evidence that a bounded context window alone cannot expose.
DIAGRAM
This diagram shows how CABLE expands host retrieval results at inference time using stored antecedent links, seed selection, scoring, and novelty filtering.
DIAGRAM
This diagram shows how CABLE is evaluated with A-MEM, SimpleMem, and Mem0g on LoCoMo and MA-LongMemEval using different LLM backbones.
PROCESS
Antecedent-Oriented Query Generation
CABLE classifies each new memory and generates a small set of antecedent-oriented queries Q(mi) to hypothesize prior experiences, causes, or motivations.
Dual Retrieval
CABLE runs Direct search to get Bi around mi and Antecedent search over Q(mi) to get Hi, broadening candidate discovery beyond direct similarity.
Overlap Subtraction
CABLE computes Ci = Hi \ Bi, explicitly removing memories already in the direct semantic neighborhood so only retriever-complementary candidates remain.
Verification and Graph Update
CABLE uses LLM verification to keep only useful antecedents, then adds directed edges mj → mi into graph G for later retrieval-time expansion.
KEY CONTRIBUTIONS
Retriever complementarity as a design principle
CABLE formalizes retriever complementarity for structured memory, using Dual Retrieval and Overlap Subtraction so links extend rather than duplicate the host retriever’s reach.
CABLE link construction and reuse
CABLE introduces Antecedent-Oriented Query Generation plus Verification and Graph Update, building sparse antecedent links once and reusing them for retrieval-time expansion without extra LLM calls.
Consistent gains across systems and backbones
CABLE improves A-MEM, SimpleMem, and Mem0g, including a +6.00 percentage point gain on MA-LongMemEval with Qwen3.5-27B and +2.20 points on LoCoMo with Mem0g.
RESULTS
Mean LLM judge LoCoMo A MEM Qwen3.5 27B
74.81%
+3.58 over A-MEM baseline 71.23%
Mean LLM judge LoCoMo A MEM DeepSeek chat
70.26%
+2.11 over A-MEM baseline 68.15%
Mean LLM judge MA LongMemEval A MEM Qwen3.5 27B
65.33%
+6.00 over A-MEM baseline 59.33%
Mean LLM judge LoCoMo Mem0g Qwen3.5 27B
54.80%
+2.20 over Mem0g baseline 52.60%
These metrics come from LoCoMo and MA-LongMemEval, which test long-term conversational memory and multi-session reasoning. The +6.00 point gain on MA-LongMemEval shows CABLE substantially improves evidence reachability in A-MEM under bounded retrieval.
BENCHMARK
These metrics come from LoCoMo and MA-LongMemEval, which test long-term conversational memory and multi-session reasoning. The +6.00 point gain on MA-LongMemEval shows CABLE substantially improves evidence reachability in A-MEM under bounded retrieval.
BENCHMARK
Mean LLM-judge score (%) for A-MEM on LoCoMo and MA-LongMemEval with different backbones.
KEY INSIGHT
On MA-LongMemEval with Qwen3.5-27B, CABLE boosts A-MEM’s mean LLM-judge score by +12.00 points on multi-session questions and +23.33 points on single-session-preference questions.
This is surprising because CABLE only replaces up to five lower-ranked entries under a fixed retrieval budget, yet the sparse antecedent links still unlock large gains where evidence is distributed across sessions.
WHY IT MATTERS
CABLE gives long-term memory systems a reusable graph of verified antecedent links, letting agents recover causes, motivations, and background events beyond direct similarity neighborhoods.
Builders can now design memory-augmented agents whose retrieval layer explicitly targets retriever-complementary evidence, improving reasoning over long, multi-session histories without increasing context size or retrieval-time LLM calls.
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…