AuthorsTejas Pawar, Sarika Patil, Om Tilekar et al.
TL;DR
IMDMR uses a six dimensional memory retrieval architecture with intelligent query processing to reach an overall score of 0.792 vs 0.207 for spaCy + RAG.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
IMDMR targets conversational systems where state of the art baselines reach only about 20% overall performance, with spaCy + RAG at 0.207.
These single dimensional RAG and memory systems miss entity relationships and temporal context, so long term conversations lose personalization and coherent user specific behavior.
HOW IT WORKS
IMDMR centers on a Memory Storage Layer, Multi-Dimensional Search Engine, Intelligent Query Processor, and Response Generation Module wired to AWS Bedrock, Amazon Titan, and Qdrant.
You can think of IMDMR like a library card catalog plus a timeline and social graph layered over vector search, instead of a single semantic index.
This architecture lets IMDMR combine six memory dimensions with a multi dimensional bonus factor, retrieving memories that a plain context window or cosine similarity store would never surface.
DIAGRAM
This diagram shows how IMDMR processes a user query through intelligent query analysis and multi dimensional search to retrieve memories and generate a response.
DIAGRAM
This diagram shows how IMDMR evaluates simulated and production variants against baselines and ablation configurations on the synthetic conversation dataset.
PROCESS
System Architecture
IMDMR initializes the Memory Storage Layer with semantic, entity, category, intent, context, and temporal metadata using Amazon Titan and Qdrant for storage.
Multi Dimensional Search
When a query arrives, IMDMR routes it to the Multi-Dimensional Search Engine, which computes dimension specific scores and applies the multi dimensional bonus factor Smulti.
Intelligent Query Processing
The Intelligent Query Processor calls AWS Bedrock for entity extraction and intent classification, then selects dimension weights and strategies based on query type.
Entity Extraction and Resolution
IMDMR runs its entity extraction and cross memory entity resolution pipeline, building entity graphs that the Response Generation Module uses to synthesize coherent answers.
KEY CONTRIBUTIONS
Multi dimensional memory retrieval architecture
IMDMR introduces a six dimension Multi-Dimensional Search Engine over a Memory Storage Layer, combining semantic, entity, category, intent, context, and temporal scores with a multi dimensional bonus factor Bmulti up to 3.0.
Intelligent query processing system
IMDMR adds an Intelligent Query Processor that uses AWS Bedrock for entity and intent analysis, dynamically selecting dimension subsets and weights for each query.
Simulation vs production comparison
IMDMR provides both IMDMR-Sim and IMDMR-Prod, showing overall scores of 0.314 and 0.792 respectively, quantifying the impact of real AWS Bedrock, Titan, and Qdrant integration.
RESULTS
Overall Score
0.792
+0.585 over spaCy + RAG
Entity Extraction F1 Score
1.000
vs spaCy + RAG at 0.500
Memory Relevance
1.000
vs spaCy + RAG at 0.333
BLEU Score
0.800
vs spaCy + RAG at 0.058
On the synthetic 1,000 conversation dataset testing entity extraction, intent understanding, and answer quality, IMDMR-Prod achieves an overall score of 0.792 compared to 0.207 for spaCy + RAG, proving that multi dimensional retrieval plus real AWS integration drastically improves conversational memory systems.
BENCHMARK
On the synthetic 1,000 conversation dataset testing entity extraction, intent understanding, and answer quality, IMDMR-Prod achieves an overall score of 0.792 compared to 0.207 for spaCy + RAG, proving that multi dimensional retrieval plus real AWS integration drastically improves conversational memory systems.
BENCHMARK
Overall Score across IMDMR variants and baseline systems.
BENCHMARK
Overall Score for IMDMR_Full vs single and hybrid dimension variants.
KEY INSIGHT
IMDMR shows that IMDMR-Prod reaches a perfect entity extraction F1 of 1.000 while IMDMR-Sim only reaches 0.667 and spaCy + RAG 0.500.
This is surprising because many assume simulated pipelines are enough for memory research, but IMDMR reveals real AWS Bedrock and Titan integration can add 0.333 F1 over a sophisticated simulated system.
WHY IT MATTERS
IMDMR unlocks a practical way to treat conversational memory as a six dimensional object, not just a semantic vector, with explicit entity, category, intent, context, and temporal axes.
With IMDMR, builders can deploy production agents that remember user preferences and goals over time, achieving 0.792 overall scores where classic RAG stacks like spaCy + RAG plateau around 0.207.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…