Cognis: Context-Aware Memory for Conversational AI Agents

Parshva Daftari; Khush Patel; Shreyas Kapale; Jithin George; Siva Surendira

cs.CLMar 27, 2026

Cognis: Context-Aware Memory for Conversational AI Agents

Authors: Parshva Daftari, Khush Patel, Shreyas Kapale, Jithin George, Siva Surendira

Organizations: Lyzr Research

Abstract

LLM agents lack persistent memory, causing conversations to reset each session and preventing personalization over time. We present Lyzr Cognis, a unified memory architecture for conversational AI agents that addresses this limitation through a multi-stage retrieval pipeline. Cognis combines a dual-store backend pairing OpenSearch BM25 keyword matching with Matryoshka vector similarity search, fused via Reciprocal Rank Fusion. Its context-aware ingestion pipeline retrieves existing memories before extraction, enabling intelligent version tracking that preserves full memory history while keeping the store consistent. Temporal boosting enhances time-sensitive queries, and a BGE-2 cross-encoder reranker refines final result quality. We evaluate Cognis on two independent benchmarks -- LoCoMo and LongMemEval -- across eight answer generation models, demonstrating state-of-the-art performance on both. The system is open-source and deployed in production serving conversational AI applications.

Figures & tables

Panel of Figure 1 — Figure 1: LoCoMo benchmark F1 scores across four question types. Cognis achieves the highest F1 in every category. Percentage labels show Cognis’s gain over the strongest baseline. Mem0 = Mem0 retrieval; Zep = Zep memory framework; Mem0g = Mem0 with graph memory enabled; Cognis = Lyzr Cognis (ours).

Panel of Figure 2 — Figure 2: Cross-system accuracy on LongMemEval across six question types. Cognis (orange) leads on five of six categories, with the largest gains on preference recall (+33.3%) and temporal reasoning (+20.6%). SuperMemory leads only on SS-Assistant. Zep/Graphiti = Zep with Graphiti; SuperMemory = SuperMemory graph memory; Cognis = Lyzr Cognis (ours, best across answer models).

Store	Data Type	Purpose
OpenSearch	Documents	Primary store for memories and messages, native BM25 search with configurable text analysis, version history tracking
VDB	Vectors (768D + 256D)	Matryoshka semantic search with dual-dimension collections

Table 1: Dual-Store Data Distribution

Collection	Dimension	Purpose
*_768d	768D	Memory embeddings for accurate semantic retrieval
*_256d	256D	Memory embeddings for fast shortlisting (200 candidates)
immediate_recall_256d	256D	Raw message embeddings for immediate recall queries

Table 2: VDB Collection Structure

Panel of Figure 3 — Figure 3: Dual-store architecture: OpenSearch (documents + native BM25) and VDB (Matryoshka embeddings at 768D/256D).

Panel of Figure 4 — Figure 4: Two-panel ingestion pipeline. Left : Immediate storage enables recall before extraction. Right : Context-aware extraction retrieves similar memories, LLM decides ADD/UPDATE/DELETE, version tracking maintains history.

Panel of Figure 5 — Figure 5: Context-aware extraction comparison. Left : Without context, the LLM creates conflicting memories (existing memory ignored). Right : With context retrieval, the LLM issues UPDATE operations maintaining consistency.

Operation	When Applied
ADD	New information not present in existing memories
UPDATE	New information supersedes or refines an existing memory (includes replaces_id to link versions)
DELETE	New information explicitly contradicts or invalidates an existing memory
NONE	Information is a duplicate or not worth storing (e.g., small talk)

Table 3: LLM Operation Decision Logic

Panel of Figure 6 — Figure 6: Version chaining for memory history. Each UPDATE creates a new version linked via replaces_id . Historical versions have is_current=false , enabling time-travel queries that traverse the chain to retrieve complete evolution of facts.

Panel of Figure 7 — Figure 7: Retrieval pipeline: Query analysis $\rightarrow$ parallel Vector/BM25 search $\rightarrow$ RRF fusion (70%+30%) $\rightarrow$ temporal boost, dedup $\rightarrow$ BGE-2 rerank $\rightarrow$ results.

Panel of Figure 8 — Figure 8: Matryoshka two-stage retrieval: truncated 256D embeddings enable fast shortlisting (200 candidates in $\sim$ 5ms), followed by accurate 768D re-ranking ( $\sim$ 15ms) for final results.

	Single-Hop			Multi-Hop			Open-Domain			Temporal
Method	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$
LoCoMo	25.02	19.75	–	12.04	11.16	–	40.36	29.05	–	18.41	14.77	–
ReadAgent	9.15	6.48	–	5.31	5.12	–	9.67	7.66	–	12.60	8.87	–
MemoryBank	5.00	4.77	–	5.56	5.94	–	6.61	5.16	–	9.68	6.99	–
MemGPT	26.65	17.72	–	9.15	7.44	–	41.04	34.34	–	25.52	19.44	–
A-Mem	27.02	20.09	–	12.14	12.00	–	44.65	37.06	–	45.85	36.67	–

Table 4: Performance comparison on LoCoMo benchmark across question types. Metrics: F1 score, BLEU-1 (B1), LLM-as-a-Judge (J). Best in bold , second-best underlined . (

\uparrow

) = higher is better.

	Single-Hop			Multi-Hop			Open-Domain			Temporal
Embedding Model	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$
Nomic Embed	50.23	41.26	69.15	29.02	21.17	52.08	50.72	41.48	68.61	60.48	52.69	72.59
Gemma Embed	40.94	30.88	71.88	23.22	14.72	69.23	44.85	33.37	58.33	54.01	42.46	70.27
Nomic V2	42.38	31.42	65.96	30.31	23.71	50.00	51.67	42.44	68.12	58.54	51.23	73.21
Jina V3	41.42	30.10	65.25	30.41	22.81	52.08	48.66	40.19	67.18	59.37	51.85	74.77

Table 5: Embedding model ablation on LoCoMo benchmark. All configurations use the same retrieval pipeline (RRF fusion + BGE-2 reranking + OpenSearch BM25). Best per-column in bold .

	Single-Hop			Multi-Hop			Open-Domain			Temporal
Configuration	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$
BGE-2 Reranker (MongoDB)	45.33	33.96	70.21	30.55	23.69	51.04	53.76	44.65	71.34	60.86	53.43	76.64
Zero Entropy Reranker	45.95	34.86	68.23	27.50	20.75	46.97	52.11	43.98	69.59	59.73	58.76	76.17
LLM Query Decomposition	44.73	34.86	69.52	29.50	21.09	50.79	46.11	37.73	68.45	60.75	57.65	77.15
BGE-2 + OpenSearch	48.66	35.76	71.99	31.51	23.22	51.04	54.77	45.12	85.85	62.68	58.95	77.26

Table 6: Retrieval pipeline ablation on LoCoMo benchmark. All configurations use the same embedding model. Best per-column in bold .

	Single-Hop			Multi-Hop			Open-Domain			Temporal
Config (Vec/BM25)	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$	F1 $\uparrow$	B1 $\uparrow$	J $\uparrow$
70/30	48.66	35.76	71.99	31.51	23.22	51.04	54.77	45.12	85.85	62.68	58.95	77.26
75/25	49.94	36.84	72.53	30.82	22.43	50.17	55.21	46.92	86.11	62.11	58.84	77.26
80/20	49.82	36.41	72.08	29.87	22.04	49.62	55.34	46.08	86.74	61.21	57.33	74.94
60/40	49.11	35.94	71.88	28.63	20.71	48.37	55.02	45.96	86.28	60.73	55.82	74.41
50/50	49.44	36.08	70.11	27.18	19.44	45.72	52.18	40.63	83.02	60.58	55.91	74.08

Table 7: RRF weight ablation on LoCoMo benchmark. Seven vector/BM25 weight configurations evaluated with all other pipeline components held constant. Best per-column in bold .

Question Type	Accuracy	Hit@K	MRR	NDCG
SS-User	100.0%	0.91	0.88	0.87
SS-Preference	93.3%	0.97	0.64	0.67
Knowledge Update	92.3%	0.94	0.75	0.79
SS-Assistant	87.5%	0.96	0.88	0.89
Multi-Session	86.5%	0.68	0.43	0.48
Temporal Reasoning	84.2%	0.80	0.53	0.57

Table 8: LongMemEval benchmark results by question type (GPT-4.1 judge). Accuracy measures end-to-end correctness; retrieval metrics assess whether relevant memories were retrieved. Best accuracy in bold .

Question Type	Zep/Graphiti	SuperMemory	Cognis
SS-User	92.9	97.1	100.0
SS-Assistant	80.4	96.4	92.9
SS-Preference	56.7	70.0	93.3
Knowledge Update	83.3	88.5	96.2
Temporal Reasoning	62.4	76.7	92.5
Multi-Session	57.9	71.4	87.2

Table 9: Cross-system accuracy (%) on LongMemEval. Cognis column shows the best result across all eight answer models for each question type (GPT-4.1 judge). Best per-row in bold .

Answer Model	SS-User	SS-Asst	SS-Pref	Know. Upd	Temporal	Multi-Sess	OVERALL
Gemini 3 Flash	97.1	80.4	50.0	92.3	85.0	78.2	83.4
GPT-4o	97.1	71.4	93.3	93.6	85.0	78.2	85.2
Claude Haiku 4.5	95.7	85.7	80.0	85.9	85.7	80.5	85.4
GPT-4.1	100.0	87.5	93.3	93.6	82.7	79.7	87.2
GPT-5-mini	98.6	87.5	90.0	89.7	90.2	83.5	89.2
Claude Sonnet 4.6	100.0	89.3	86.7	93.6	88.0	85.7	90.0

Table 10: Cognis accuracy (%) on LongMemEval across answer generation models (GPT-4.1 judge). Best per-column in bold .

Method	p50	p95	p99	Mean
Cognis	0.250s	0.451s	0.770s	0.390s

Table 11: Cognis end-to-end retrieval latency (500 LongMemEval queries).

Agent Memory Management

Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.

See all 446 papers

MemMachine: A Ground-Truth-Preserving Memory System for Personalized AI Agents
Apr 6, 2026Shu Wang, Edwin Yu, Oscar Love +4LLM Agent Memory Memory Management for LLM Agents
MemVerge, Inc.
HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents
Apr 20, 2026Shuqi Cao, Jingyi He, Fei TanPersistent Memory for Language Models Hierarchical Memory
East China Normal University, Shanghai, China · Shanghai Jiao Tong University, Shanghai, China
Memori: A Persistent Memory Layer for Efficient, Context-Aware LLM Agents
Mar 20, 2026Luiz C. Borro, Luiz A. B. Macarini, Gordon Tindall +2Persistent Memory for Language Models LLM Memory
Memori Labs Inc.

Apr 6, 2026J/K move · Enter open · S save

MemMachine: A Ground-Truth-Preserving Memory System for Personalized AI Agents

Shu Wang, Edwin Yu, Oscar Love, Tom Zhang+3

MemVerge, Inc.

LLM Agent Memory Memory Management for LLM Agents Persistent Memory for Language Models Long-Term Conversational Memory Agent Memory Management Agentic RAG

Open paper AskPDF arXiv

Large Language Model (LLM) agents require persistent memory to maintain personalization, factual continuity, and long-horizon reasoning, yet standard context-window and retrieval-augmented generation (RAG) pipelines degrade over multi-session interactions. We present MemMachine, an open-source memory system that integrates short-term, long-term episodic, and profile memory within a ground-truth-preserving architecture that stores entire conversational episodes and reduces lossy LLM-based extraction. MemMachine uses contextualized retrieval that expands nucleus matches with surrounding context, improving recall when relevant evidence spans multiple dialogue turns. Across benchmarks, MemMachine achieves strong accuracy-efficiency tradeoffs: on LoCoMo it reaches 0.9169 using gpt4.1-mini; on LongMemEvalS (ICLR 2025), a six-dimension ablation yields 93.0 percent accuracy, with retrieval-stage optimizations -- retrieval depth tuning (+4.2 percent), context formatting (+2.0 percent), search prompt design (+1.8 percent), and query bias correction (+1.4 percent) -- outperforming ingestion-stage gains such as sentence chunking (+0.8 percent). GPT-5-mini exceeds GPT-5 by 2.6 percent when paired with optimized prompts, making it the most cost-efficient setup. Compared to Mem0, MemMachine uses roughly 80 percent fewer input tokens under matched conditions. A companion Retrieval Agent adaptively routes queries among direct retrieval, parallel decomposition, or iterative chain-of-query strategies, achieving 93.2 percent on HotpotQA-hard and 92.6 percent on WikiMultiHop under randomized-noise conditions. These results show that preserving episodic ground truth while layering adaptive retrieval yields robust, efficient long-term memory for personalized LLM agents.

Abstract

Agent Memory Management

Explore similar work

MemMachine: A Ground-Truth-Preserving Memory System for Personalized AI Agents

HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents

Memori: A Persistent Memory Layer for Efficient, Context-Aware LLM Agents