APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI

Pratyay Banerjee; Masud Moshtaghi; Shivashankar Subramanian; Amita Misra; Ankit Chadha

cs.CLApr 15, 2026

APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI

Authors: Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian, Amita Misra, Ankit Chadha

Organizations: Amazon, AGI, Sunnyvale, USA

Abstract

Large language models still struggle with reliable long-term conversational memory: simply enlarging context windows or applying naive retrieval often introduces noise and destabilizes responses. We present APEX-MEM, a conversational memory system that combines three key innovations: (1) a property graph which uses domain-agnostic ontology to structure conversations as temporally grounded events in an entity-centric framework, (2) append-only storage that preserves the full temporal evolution of information, and (3) a multi-tool retrieval agent that understands and resolves conflicting or evolving information at query time, producing a compact and contextually relevant memory summary. This retrieval-time resolution preserves the full interaction history while suppressing irrelevant details. APEX-MEM achieves 88.88% accuracy on LOCOMO's Question Answering task and 86.2% on LongMemEval, outperforming state-of-the-art session-aware approaches and demonstrating that structured property graphs enable more temporally coherent long-term conversational reasoning.

Figures & tables

Figure 1: End-to-end pipeline for constructing and querying APEX-MEM Graph, showing data flow from unstructured conversation to GraphSQL QA Agent.

Method			Single-Hop	Multi-Hop	Temporal	Open-Domain	Adversarial	Overall	w/o Adv.
APEX-MEM	Graph	QA Agent
	APEX-MEM	Claude 4.5 Haiku	85.46%	84.74%	79.17%	89.18%	87.22 %	84.92%	84.25%
	APEX-MEM	Claude 4.5 Sonnet	89.36 %	86.92 %	90.63 %	87.75%	86.10%	88.41%	89.08%
	APEX-MEM	Claude 3.5 Sonnet	87.58%	84.74%	87.50%	88.94%	86.10%	86.90%	87.13%
	APEX-MEM	GPT5	89.88%	86.29%	90.63 %	91.68 %	86.77%	88.88 %	89.49 %
	APEX-MEM	GPT4o	88.47%	85.46%	83.49%	86.46%	84.98%	86.35%	86.75%

Table 1: LOCOMO Category Type Evaluation Results

Model	Fact	Schema	Entity/Prop.
	Extraction	Coverage	Resolution
GPT4o	94.2%	75.7%	98.1%
Claude Sonnet 4.5	97.3%	91.1%	98.2%
Claude Haiku 4.5	95.8%	90.3%	95.4%
Qwen3-14B	95.4%	88.9%	92.5%

Table 2: APEX-MEM Construction metrics for fact extraction (precision of extracted facts), schema coverage (plausible properties covered), and entity resolution (detecting and linking to proper entity). We measure these metrics with 500 Random turns from LoCoMo and LongMemEval. GPT5 is used as the judge.

Method		Single-Hop	Multi-Hop	Temporal	Open-Domain	Adversarial	Overall
Graph	QA Agent: Claude 4.5 Haiku w/ Tools
APEX-MEM	SchemaViewer, EntityLookUp	80.85%	76.64%	72.92%	76.34%	77.80%	77.19%
	+ GraphSQL	80.78%	79.75%	82.29%	78.00%	81.16%	79.45%
	+ Search	85.46%	84.74%	79.17%	89.18%	87.22%	87.00%

Table 3: APEX-MEM Ablations of different tools

Method			Overall Score
APEX-MEM	Graph	QA Agent
	APEX-MEM Online	Claude 4.5 Haiku	82.8%
		Claude 4.5 Sonnet	86.2%
		Claude 4 Sonnet	81.0%
		GPT5	85.2%
		GPT4o	75.0%

Table 4: LongMemEval Evaluation Results

Method		Acc.
APEX-MEM (Online)	QnAAgent
	Claude 4 Sonnet	28.9%
	Claude 4.5 Sonnet	35.2%
	GPT5	40.1 %
	GPT4o	19.0%
Baselines w/ Web-Search Tool

Table 5: SealQA Evaluation Results

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Tool	Claude 4.5 Haiku	Claude 4.5 Haiku: Entity	Claude 4.5 Haiku: GraphSQL	Claude 4.5 Sonnet
EntityLookup	5160	17414	5546	4894
Search	8900	1058	0	10544
PropertySearch	3346	358	5016	3342
SchemaViewer	3958	4004	3622	3964
GraphSQL	8260	0	27282	7168

Appendix

Table 6: Tool Call Distribution Across Methods

SQL Category	Claude 4.5 Haiku	Claude 4.5 Haiku: Entity	Claude 4.5 Haiku: GraphSQL	Claude 4.5 Sonnet
SELECT	315	0	1850	88
JOIN	81	0	430	31
AGGREGATE	1417	0	4586	1230
TEMPORAL	2317	0	6775	2235

Appendix

Table 7: SQL Query Categories by Method

Category	Example Query
SELECT	SELECT DISTINCT entity_id, entity_name, entity_type FROM entities WHERE entity_name LIKE ’%Anthony%’ COLLATE NOCASE
JOIN	SELECT f.property_name, f.value_json, f.dtype FROM facts f JOIN evidence e ON e.fact_id = f.id WHERE e.event_id = 518
AGGREGATE	SELECT COUNT(DISTINCT device_name) as device_count FROM ( SELECT ’Fitbit Versa 3’ as device_name UNION SELECT ’nebulizer machine’ as device_name UNION SELECT ’Accu-Chek Aviva Nano’ as device_name UNION SELECT ’hearing aids’ as device_name )
TEMPORAL	SELECT f.value_json as start_date, date(:question_date) as question_date, julianday(:question_date) - julianday(json_extract(f.value_json, ’ $’)) as days_difference, CAST((julianday(:question_date) - julianday(json_extract(f.value_json, ’$ ’)))/ 30.44 AS INTEGER) as months_approx, CAST((julianday(:question_date) - julianday(json_extract(f.value_json, ’$’)))/7 AS INTEGER) as weeks_approx FROM facts f WHERE f.subject_id = :user_id AND f.property_name = :property_id ORDER BY f.created_at DESC LIMIT 1

Appendix

Table 8: Sample SQL Queries by Category

Figure 2: Analysis of Tool Calls v/s Accuracy on LOCOMO Dataset. We cap the max Tool calls at 40.

Method	GC	GC	Amort.	Mem	QnA	Total	Acc
	Tok/Conv	Calls	GC/Q	Tok/Q	Tok/Q	Tok/Q	(w/o Adv)
MIRIX (est.)	$\sim$ 15.2M	$\sim$ 4,704	$\sim$ 98,750	$\sim$ 4,500	$\sim$ 13,500	$\sim$ 112,000	85.38%
Zep/Graphiti	$\sim$ 9.4M	$\sim$ 5,292	$\sim$ 60,900	2,247	$\sim$ 3,900	$\sim$ 64,800	75.14%
Mem0 g (est.)	$\sim$ 4.9M	$\sim$ 2,352	$\sim$ 31,882	3,616	$\sim$ 5,300	$\sim$ 37,200	68.44%
APEX-MEM	2.69M	3,717	13,557	8,000 ‡	$\sim$ 16,000	$\sim$ 30,000	84–89% †
Full Context	0	0	0	23,653	$\sim$ 25,000	$\sim$ 25,000	87.52%

Appendix

Table 9: Token consumption comparison across memory methods. † Accuracy depends on QnA model. ‡ Tunable via top-

k

retrieval budget. GC = Graph/Memory Construction; Amort. = Amortized over queries per conversation.

Component	Tok/Q	%	Source
Graph constr. (amort.)	13,557	16.6	Extraction, resolution
System prompt	7,854	9.6	Fixed per architecture
Memory retrieval	21,745	26.6	Graph search (top- $k$ )
Agent loop overhead	16,174	19.8	Prior msgs in loop
Tool framing	22,274	27.3	Tool call/response fmt
Total (mean)	81,604	100

Appendix

Table 10: APEX-MEM token decomposition per query.

Metric	Sonnet	GPT-5	Sonnet	Haiku
	4.5		(SQL)
SQL Executions	3,659	2,163	66,580	4,277
Successful	3,574	2,020	65,900	4,080
SQL Errors	87	143	791	197
Success Rate	97.6%	93.4%	98.8%	95.4%

Appendix

Table 11: GraphSQL execution analysis across LLM backends. “Sonnet (SQL)” denotes the GraphSQL-only ablation.

System	Append-Only?	Temporal	$\Delta$
APEX-MEM	Yes	90.63%	—
Mem0	No (consolid.)	75.71%	–14.92
MIRIX	No (state merge)	65.62%	–25.01
Zep	Partial (temp. KG)	76.60%	–14.03

Appendix

Table 12: LOCOMO temporal accuracy: append-only vs. eager update strategies.

\Delta

is the difference from APEX-MEM in percentage points.

Original Text	Entities	Types	Properties	Values	Data Types
"I went to a LGBTQ support group yesterday and it was so powerful."	Caroline	PERSON	P:attended_event	"LGBTQ support group"	str
			P:experience_ description	"powerful"	str
	LGBTQ support group	EVENT	P:event_date	"2023-05-07"	date
			P:event_type	"support group"	str
"I just signed up for a pottery class yesterday. It’s like therapy for me."	Melanie	PERSON	P:activities_ participated	"pottery class"	str
			P:activity_benefit	"therapy for me"	str

Appendix

Table 13: Examples of text-to-graph conversion in APEX-MEM. Each conversational utterance is systematically transformed into structured entity-property-value triplets with appropriate semantic types, enabling sophisticated querying and temporal reasoning over conversational knowledge.

Long-Term Conversational Memory

Long-term conversational memory helps AI assistants retain, update, and use information from past conversations across sessions.

See all 255 papers

EngramaBench: Evaluating Long-Term Conversational Memory with Structured Graph Retrieval
Apr 23, 2026Julian AcunaLLM Evaluation Memorization in Language Models
JustMem: Just-Enough Memory Access for Long-Term Conversations
Sep 17, 2026Guanhua Chen, Yanting Wang, Wenjing Zhi +1Conversational Memory Persistent Memory for Language Models
Beihang University
GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory
May 3, 2026Yushi Sun, Bowen Cao, Dong Fang +2LLM Agent Memory Persistent Memory for Language Models
LIGHTSPEED, Shenzhen, China · The Chinese University of Hong Kong, Hong Kong, China

Apr 23, 2026J/K move · Enter open · S save

EngramaBench: Evaluating Long-Term Conversational Memory with Structured Graph Retrieval

Julian Acuna

LLM Evaluation Memorization in Language Models Persistent Memory for Language Models Memorization in Generative Models Memory Agent Benchmarks Graph-Based Agent Memory+1

Open paper AskPDF arXiv

Large language model assistants are increasingly expected to retain and reason over information accumulated across many sessions. We introduce EngramaBench, a benchmark for long-term conversational memory built around five personas, one hundred multi-session conversations, and one hundred fifty queries spanning factual recall, cross-space integration, temporal reasoning, adversarial abstention, and emergent synthesis. We evaluate Engrama, a graph-structured memory system, against GPT-4o full-context prompting and Mem0, an open-source vector-retrieval memory system. All three use the same answering model (GPT-4o), isolating the effect of memory architecture. GPT-4o full-context achieves the highest composite score (0.6186), while Engrama scores 0.5367 globally but is the only system to score higher than full-context prompting on cross-space reasoning (0.6532 vs. 0.6291, n=30). Mem0 is cheapest but substantially weaker (0.4809). Ablations reveal that the components driving Engrama's cross-space advantage trade off against global composite score, exposing a systems-level tension between structured memory specialization and aggregate optimization.

Abstract

Long-Term Conversational Memory

Explore similar work

EngramaBench: Evaluating Long-Term Conversational Memory with Structured Graph Retrieval

JustMem: Just-Enough Memory Access for Long-Term Conversations

GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory