AuthorsSamarth Sarin, Lovepreet Singh, Bhaskarjit Sarmah, Dhagash Mehta
TL;DR
Memoria uses weighted knowledge graph triplets plus session-level summarization to deliver 87.1% accuracy on LongMemEvals single-session-user questions, beating A-Mem (OpenAI) by 2.9 points.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Memoria targets LLM systems where each interaction is treated in isolation, discarding previous context and failing to adapt over time.
This breaks personalized conversational assistants and agentic memory use cases, causing repetitive questions, weak long term context, and lower user trust.
HOW IT WORKS
Memoria’s core mechanism combines structured conversation logging, dynamic user persona via KG, session level memory for real time context, and seamless retrieval for context aware responses.
You can think of Memoria like RAM plus disk: session summaries act as fast working context, while the weighted KG behaves like a long term card catalog of user facts.
This design lets Memoria inject only the most relevant, recency weighted triplets and summaries into the prompt, enabling continuity that a plain context window cannot provide.
DIAGRAM
This diagram shows how Memoria routes new and repeat users through session summaries and KG retrieval to build prompts.
DIAGRAM
This diagram shows how Memoria and A-Mem are evaluated on LongMemEvals with shared LLM and different embedding setups.
PROCESS
Structured Conversation History with Database
Memoria logs each user query and LLM reply with timestamps, session identifiers, KG triplets, summaries, and token usage into the structured conversation logging module.
Dynamic User Persona via KG
Memoria extracts subject predicate object triplets from user messages, embeds them, and grows a weighted knowledge graph capturing recurring topics, preferences, and entities.
Session Level Memory for Real Time Context
Memoria incrementally builds session level memory for real time context by summarizing dialogue turns and storing summaries keyed by session identifiers.
Seamless Retrieval for Context Aware Responses
Memoria performs seamless retrieval for context aware responses by fetching session summaries and top K KG triplets, applying exponential decay weights, and constructing a distilled prompt.
KEY CONTRIBUTIONS
Structured Conversation History with Database
Memoria introduces a structured conversation logging layer that stores raw messages, KG triplets, summaries, and token statistics in SQLite3 plus ChromaDB, turning chats into a queryable memory bank.
Dynamic User Persona via KG
Memoria builds a dynamic user persona via KG by embedding and linking user specific triplets, then weighting them with an exponential decay function to resolve conflicting preferences over time.
Session Level Memory and Seamless Retrieval
Memoria combines session level memory for real time context with seamless retrieval for context aware responses, achieving 87.1% single session accuracy and up to 38.7% latency reduction versus full context prompting.
RESULTS
Single Session accuracy
87.1%
+2.9 over A-Mem (OA)
Knowledge Update accuracy
80.8%
+1.4 over A-Mem (OA)
Inference Time single session-user
260 secs
131 secs faster than Full Context
Avg Token length single session-user
398 tokens
vs 115000 tokens for Full Context
On LongMemEvals single-session-user and knowledge-update subsets, which stress user specific recall and knowledge updates, Memoria shows that recency weighted KG memory can match or exceed full context accuracy with far shorter prompts.
BENCHMARK
On LongMemEvals single-session-user and knowledge-update subsets, which stress user specific recall and knowledge updates, Memoria shows that recency weighted KG memory can match or exceed full context accuracy with far shorter prompts.
BENCHMARK
Accuracy on LongMemEvals single-session-user questions.
KEY INSIGHT
Memoria with weighted KG and summaries reaches 87.1% single session accuracy, slightly above the 85.7% full conversation baseline while using tiny prompts.
This is surprising because many assume feeding the entire 115000 token history is always best, yet Memoria’s curated memory beats that with only about 398 tokens.
WHY IT MATTERS
Memoria unlocks practical agentic memory by combining structured logging, KG based personas, and recency aware retrieval into a plug and play Python framework.
Builders can now ship conversational agents that remember users across long horizons without exploding token budgets, enabling persistent assistants, advisors, and support bots on modest infrastructure.
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.
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.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…