Memoria: A Scalable Agentic Memory Framework for Personalized Conversational AI (2512.12686v1)

Abstract: Agentic memory is emerging as a key enabler for LLMs (LLM) to maintain continuity, personalization, and long-term context in extended user interactions, critical capabilities for deploying LLMs as truly interactive and adaptive agents. Agentic memory refers to the memory that provides an LLM with agent-like persistence: the ability to retain and act upon information across conversations, similar to how a human would. We present Memoria, a modular memory framework that augments LLM-based conversational systems with persistent, interpretable, and context-rich memory. Memoria integrates two complementary components: dynamic session-level summarization and a weighted knowledge graph (KG)-based user modelling engine that incrementally captures user traits, preferences, and behavioral patterns as structured entities and relationships. This hybrid architecture enables both short-term dialogue coherence and long-term personalization while operating within the token constraints of modern LLMs. We demonstrate how Memoria enables scalable, personalized conversational AI by bridging the gap between stateless LLM interfaces and agentic memory systems, offering a practical solution for industry applications requiring adaptive and evolving user experiences.

• The paper introduces a hybrid architecture that integrates session summaries with a weighted knowledge graph to enable personalized, long-term conversational memory.
• It employs an Exponential Weighted Average method for conflict resolution and efficient context retrieval, achieving up to 87.1% accuracy on test benchmarks.
• The approach reduces inference latency by 38.7% and token usage while allowing plug-and-play integration with existing LLM systems.

Memoria: A Scalable Agentic Memory Framework for Personalized Conversational AI

Overview and Motivation

The "Memoria" framework addresses the persistent challenge in LLM-based conversational AI: the lack of agentic memory that supports long-term context, user personalization, and continuity across sessions. Stateless LLM deployments inherently treat each interaction as an isolated event, discarding valuable contextual signals and thereby limiting coherence and adaptive potential. To overcome these restrictions, Memoria introduces a hybrid system that fuses dynamic session-level summarization with an incrementally built, weighted user-specific knowledge graph (KG), aiming to achieve scalable, interpretable, and richly personalized dialogue.

System Architecture

The architectural design of Memoria supports plug-and-play integration with any LLM-based chat system. Four synergistic core modules underpin the system: (1) structured conversation storage, (2) session-level summarization, (3) knowledge graph-based user modeling, and (4) context-aware retrieval with recency weighting. This enables Memoria to preserve both short-term and long-term user data in a manner that is compatible with LLM token constraints and memory limits.

Figure 1: Memoria architecture combines persistent conversation storage, session summarization, and a weighted KG user model for context-rich, adaptive dialogue generation.

The architecture distinguishes between various user-session scenarios, optimizing what information is incorporated into each prompt and continuously updating memory modules after every exchange. When a new user initiates a session, minimal context is available; as the system accumulates user data, relevant triplets and session summaries are seamlessly injected into future prompts.

Operational Scenarios

Memoria's utility is exemplified across three principal user interaction patterns:

New User with New Session: The system operates statelessly on first contact, storing initial exchanges and extracting raw knowledge triplets from the user’s message to initialize their personal KG. Session and user summaries are subsequently created for downstream personalization.

Figure 2: Information flow for a new user and new session—minimal context, bootstrapping of persistent storage.

Repeat User with New Session: The system identifies returning users by referencing stored KG triplets—enabling session personalization with previously inferred preferences and traits, despite a new conversation thread. Prompt augmentation leverages these attributes from the initial turn.

Figure 3: Repeat user with a new session utilizes the evolving KG for immediate personalization, even when the session context is new.

Repeat User with Repeat Session: Both session and user memory are available from the outset. Session summaries and top-weighted, semantically filtered KG triplets are merged into the prompt, enabling real-time, highly context-aware dialogue.

Figure 4: For users continuing an active session, Memoria supplies the LLM with both long-term (KG) and short-term (session summary) memory for deep continuity.

Memory Update and Retrieval Strategy

Memoria’s dual-memory update mechanism enables continuous, bidirectional enrichment of the user representation. Summarization of each conversational turn ensures that the session-level summary consistently reflects the current state. For the KG, each new user message is parsed into knowledge triplets, which are persisted with associated metadata (semantics, timestamp, and session information). These triplets are embedded and stored in a vector database for efficient semantic retrieval.

A key innovation of Memoria is its application of an Exponential Weighted Average (EWA) decay to KG triplets. This weighting policy ensures higher priority for recent user inputs, facilitating effective conflict resolution (for instance, where a user's new statement contradicts prior claims) and aligning LLM responses with the latest user intent. The weighting is governed by an exponential decay of the form:

$w_i = \frac{e^{-a x_i}}{\sum_{j=1}^N e^{-a x_j}}$

where $x_i$ represents the (normalized) elapsed time since triplet creation, $a$ is the decay rate, and $N$ is the size of the retrieved triplet set. This prioritizes recency while maintaining a soft memory of distant interactions when relevant.

Experimental Evaluation

Memoria’s efficacy is validated using the LongMemEvals benchmark, focusing on two categories: single-session-user and knowledge-update, both central to evaluating long-term factual recall and state revision in dialogue agents. Comparative experiments were conducted against A‑Mem, both in its default configuration (SentenceTransformers all-MiniLM-L6-v2 embedder) and a variant using OpenAI’s ada-002 embeddings (to match Memoria’s embedding backbone).

Memoria consistently outperforms A‑Mem in both accuracy and token efficiency.

Task	Full Context	A-Mem (ST)	A-Mem (OA)	Memoria
Single-Session User	85.7%	78.5%	84.2%	87.1%
Knowledge-Update	78.2%	76.2%	79.4%	80.8%

Memoria’s use of weighted retrieval enables it to resolve contradictions and maintain up-to-date factual grounding, achieving superior coherence compared to unweighted A‑Mem.

Latency and efficiency experiments further demonstrate a 38.7% reduction in inference time (single-session-user) versus full-context prompting, and a compression of prompt token size from 115,000 to <400 tokens. This delivers substantial computational and economic savings while preserving or exceeding answer fidelity.

Theoretical and Practical Implications

Memoria’s hybrid approach concretizes several advances in scalable agentic memory:

• Structured, Interpretable Memory: The weighted KG enables interpretable reasoning and symbolic manipulation, unlike pure vector/RAG approaches.
• Recency-aware Conflict Resolution: EWA weighting automatically resolves contradictory information in favor of the latest user input without explicit conflict heuristics.
• Plug-and-Play Integration: No LLM fine-tuning is required; all enhancements are externalized, supporting modular extension of existing LLM APIs without performance degradation.
• Efficient Context Management: The system remains compatible with strict LLM token limits by curating succinct, high-value prompt augmentations.

From a theoretical standpoint, Memoria’s architecture can be extended to support multi-agent temporal reasoning, multi-persona modeling, and domain-specific KG structuring. Its modularity suggests immediate application in areas requiring robust long-term personalization—such as financial advisory, customer service automation, and knowledge management platforms.

Future Prospects

Open research directions include:

• Extension to reasoning tasks requiring multi-modal and multi-agent memory.
• Analysis of KG quality and drift over long deployment periods.
• Integration with dynamic retrieval-augmented generation that considers not just past user state but also external evolving data sources.
• Incorporation of explicit working memory enhancement for complex reasoning workflows.

Conclusion

Memoria introduces a robust, scalable architecture for agentic memory, advancing personalized conversational AI by integrating KG-based long-term memory, session-level summarization, and recency-weighted retrieval. Its superior accuracy, token efficiency, and computational performance substantiate its value as a practical memory augmentation layer for real-world LLM deployments. The framework’s modular design and strong empirical results indicate significant potential for broader adoption and further research in agentic conversational systems (2512.12686).