AuthorsKe Yang, Zixi Chen, Xuan He et al.
TL;DR
PLUGMEM uses a knowledge-centric memory graph over propositions and prescriptions to deliver up to 1.4e-1 bits per token on HotpotQA, far higher than retrieval baselines.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long-term memory for LLM agents is either task-specific and non-transferable, or task-agnostic but less effective due to low task-relevance and context explosion from raw memory retrieval.
Vanilla retrieval-based systems that store all past interactions as raw text chunks suffer from unbounded memory growth, high computational cost, and degraded decision-making when deployed in complex environments.
HOW IT WORKS
PLUGMEM combines a Structuring Module, Retrieval Module, and Reasoning Module to transform episodic trajectories into semantic and procedural knowledge stored in a unified memory graph.
You can think of PLUGMEM like a brain with a hippocampus that turns raw experiences into compact facts and workflows, then a librarian that fetches only the shelves relevant to the current question.
By operating over propositional and prescriptive knowledge units instead of raw trajectories, PLUGMEM enables retrieval and reasoning that a plain context window cannot, achieving much higher information density per memory token.
DIAGRAM
This diagram shows how PLUGMEM retrieves task-relevant propositions and prescriptions via multi-hop abstraction-specific traversal over semantic and procedural graphs.
DIAGRAM
This diagram shows how PLUGMEM is evaluated across LongMemEval, HotpotQA, and WebArena with ablations on structuring, retrieval, and reasoning.
PROCESS
Structuring Module
PLUGMEM uses the Structuring Module to standardize episodic traces into tuples and extract semantic propositions and procedural prescriptions linked into memory graphs.
Semantic Memory
PLUGMEM builds the semantic graph by turning propositions and concepts into nodes with embeddings and provenance edges to episodic memory.
Procedural Memory
PLUGMEM segments trajectories by subgoal similarity, induces intent prescription pairs, and stores them in the procedural graph with return scores.
Retrieval Module
PLUGMEM runs the Retrieval Module to perform abstraction-aware multi-hop retrieval over semantic and procedural graphs before the Reasoning Module condenses results for the base agent.
KEY CONTRIBUTIONS
Design principles
PLUGMEM formalizes cognitively motivated principles for task-agnostic memory, separating episodic, semantic, and procedural memory via the Structuring Module and knowledge-centric graphs.
Evaluation framework
PLUGMEM introduces an information-theoretic measure of memory utility and efficiency, reporting global memory information density up to 1.4e-1 bits per token on HotpotQA.
General memory module
PLUGMEM provides a plugin memory system with Structuring, Retrieval, and Reasoning Modules, evaluated unchanged on LongMemEval, HotpotQA, and WebArena with higher information density than baselines.
RESULTS
EM
61.4
+9.7 EM over Vanilla Retrieval on HotpotQA
F1
74.1
+11.4 F1 vs Vanilla Retrieval on HotpotQA
Info. Density
1.4e-1 bits per token
about 11.7x higher than Vanilla Retrieval at 1.2e-2 on HotpotQA
#TokAvg.
81.6 tokens
about 8.1x fewer tokens than Vanilla Retrieval at 659.2 on HotpotQA
On HotpotQA, which tests multi-hop knowledge retrieval and reasoning, PLUGMEM delivers 61.4 EM and 74.1 F1 with only 81.6 memory tokens, showing that knowledge-centric memory graphs can provide much higher decision-relevant information per token than retrieval baselines.
BENCHMARK
On HotpotQA, which tests multi-hop knowledge retrieval and reasoning, PLUGMEM delivers 61.4 EM and 74.1 F1 with only 81.6 memory tokens, showing that knowledge-centric memory graphs can provide much higher decision-relevant information per token than retrieval baselines.
BENCHMARK
EM on HotpotQA for PLUGMEM and key memory baselines.
BENCHMARK
Accuracy on LongMemEval for PLUGMEM and representative baselines.
KEY INSIGHT
PLUGMEM reaches 61.4 EM and 74.1 F1 on HotpotQA with only 81.6 memory tokens, while Vanilla Retrieval needs 659.2 tokens for 51.7 EM and 62.7 F1.
This is surprising because task-agnostic memory modules are usually assumed to trade effectiveness for generality, yet PLUGMEM beats task-specific designs like RAPTOR and HippoRAG2 in information density.
WHY IT MATTERS
PLUGMEM shows that agents can share a single plugin memory that abstracts episodic traces into reusable knowledge, then retrieve it efficiently across heterogeneous tasks.
Builders can now attach PLUGMEM to new agents and benchmarks without redesigning memory, gaining long-term semantic and procedural recall without exploding context windows.
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…