State of the Art: LLM Memory Systems and Affective Computing¶

Analysis of existing systems, focusing on gaps in emotional memory implementation. This document motivates the positioning of our library within the current landscape.

1. LLM Memory Systems¶

1.1 MemGPT / Letta (Packer et al., 2023)¶

Paper: "MemGPT: Towards LLMs as Operating Systems" (arXiv:2310.08560) GitHub: https://github.com/cpacker/MemGPT → became https://research.memgpt.ai/

Architecture — Three OS-inspired tiers: - Core Memory (always in-context, like RAM): system persona, key user information - Recall Memory (searchable conversational history): vector database of past conversations - Archival Memory (long-term storage, like disk): unboundedly expandable memory

Key innovation: The LLM itself decides what to store, summarize, and forget via tool calls — self-management of memory.

Emotional gap: No emotional representation. The importance of a memory is determined by the LLM in an opaque way, without a structured affective dimension. There is no differentiated decay based on emotional features.

1.2 Mem0 (2024)¶

Paper: arXiv:2504.19413 GitHub: https://github.com/mem0ai/mem0 (~48K stars, $24M funding)

Approach: Universal memory layer for AI agents. Dynamically extracts salient information from conversations, consolidates it, and retrieves it.

Metrics: 26% higher accuracy, 91% lower latency, 90% token savings compared to full-context approaches.

Emotional gap: Extracts facts and preferences, not emotional states. Does not distinguish between "the user prefers coffee" and "the user was distressed when talking about their work."

1.3 Zep (2024)¶

GitHub: https://github.com/getzep/zep

Approach: Temporal knowledge graph that connects past interactions, structured datasets, and context changes. The Graphiti engine combines episodic memory (chat), semantic memory (entities), and group-level subgraphs.

Emotional gap: The knowledge graph captures semantic relationships but not affective dimensions. There is no node or edge representing the emotional state associated with an interaction.

1.4 Generative Agents (Park et al., 2023)¶

Paper: "Generative Agents: Interactive Simulacra of Human Behavior" (arXiv:2304.03442) Citations: ~3000+ — foundational paper for the field

Retrieval formula:

score = recency(t) + importance(m) + relevance(q, m)

- recency: exponential decay from the time of last access - importance: LLM evaluation on a 1–10 scale ("how important is this memory to the character?") - relevance: cosine similarity between query embedding and memory embedding

Emotional gap: importance is an opaque proxy that may capture emotional relevance, but: 1. It is unstructured — does not distinguish valence, arousal, emotion type 2. It does not modulate decay — a very important memory decays at the same rate as a trivial one 3. There is no mood-congruent retrieval — the character's mood does not weight retrieval 4. There is no reconsolidation

This paper defined the de facto standard that most subsequent systems reproduce with minor variations.

1.5 OpenMemory / Cavira (2025)¶

GitHub: https://github.com/CaviraOSS/OpenMemory Docs: https://openmemory.cavira.app/

Architecture — 5 sectors with differentiated decay: 1. Semantic (facts, concepts): standard decay 2. Episodic (specific events): moderate decay 3. Procedural (procedures, skills): slow decay (habits persist) 4. Emotional (emotional states, affective experiences): slow decay with minimum floor 5. Reflective (self-assessments, insights): very slow decay

Key innovation: The first system to include an explicit emotional sector with differentiated decay. "Reinforcement pulses" raise critical memories above the retention threshold.

Gap: Emotion as a separate sector, not as a pervasive dimension. The internal emotional representation is not documented in detail. No appraisal engine.

1.6 Memory Bear AI (2025–2026)¶

Paper: arXiv:2512.20651 (Dec 2025) + Technical Report arXiv:2603.22306 (Mar 2026)

Innovation: Emotion Memory Units (EMUs) — the first system to treat emotion as a native dimension of memory, not as a separate sector or post-hoc label.

Each EMU combines: - Semantic content - Emotional state at encoding time - Subsequent emotional impact traces

Relevance: The EMU approach is the closest to our vision. But it remains limited — it lacks the multi-layer structure (momentum, mood field, vectorial appraisal) and the multi-signal retrieval with adaptive weights.

1.7 Emotional RAG (Huang et al., 2024)¶

Paper: "Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval" (arXiv:2410.23041)

Theoretical foundation: Bower's Mood-Dependent Memory Theory (1981)

Two retrieval strategies: 1. Combination: joint weighting of semantic similarity + emotional state 2. Sequential: filtering by emotion then semantic ranking (or vice versa)

Outperforms standard RAG on InCharacter, CharacterEval, Character-LLM datasets.

Gap: Tested only in role-playing scenarios. No long-term memory management. Emotion is a single signal, not a multi-layer field.

1.8 A-Mem / Agentic Memory (2025)¶

Paper: "A-MEM: Agentic Memory for LLM Agents" (NeurIPS 2025, arXiv:2502.12110) GitHub: https://github.com/WujiangXu/A-mem

Approach: The LLM agent autonomously manages its own memory operations (what to remember, how to structure it, when to forget). Memory as a self-organizing adaptive system.

Emotional gap: Focus on self-management, not affective dimensions.

1.9 MemOS — Memory Operating System (2025)¶

GitHub: https://github.com/MemTensor/MemOS

Approach: Operating system for LLM memory with a knowledge base, memory feedback, multi-modal memory, and memory tools for agent planning.

Emotional gap: Systematic structure without an affective component.

2. MemEmo Benchmark: The Definitive Gap (Feb 2026)¶

Paper: "MemEmo: A Benchmark for Evaluating Emotional Memory in LLM Agents" (arXiv:2602.23944)

Key result: No current memory system correctly handles emotion across all phases: extraction, updating, and question-answering.

Dataset: HLME (Human-Like Memory with Emotion) — conversations with explicit and implicit emotional states, state changes, and queries requiring emotional reasoning over memory.

Evaluated tasks: 1. Emotion Extraction: Identify the emotional state in a conversation 2. Emotion Update: Update memory when emotion changes 3. Emotion QA: Answer questions requiring reasoning over emotional memory

Systems tested: All major systems (Mem0, Zep, MemGPT/Letta, others) fail systematically, especially on emotional memory updating and reasoning.

This paper provides strong motivation for a library that treats emotion as a structural part of memory, but does not by itself constitute a validation of our implementation.

3. Affective Computing: Contemporary Landscape¶

Multimodal Fusion¶

The dominant trend (2022–2026) combines: - Facial expression - Speech prosody - Text - Physiological signals (EDA, heart rate, EEG) - Eye tracking

Cross-modal transformers (e.g. Joint-Dimension-Aware Transformer) dominate. Accuracy > 90% in some paradigms.

Relevance: Our Layer 4 (Appraisal) can potentially accept multimodal input beyond text.

Emotional Support Conversation (ESC)¶

LLM-based systems for emotional support: they detect the user's emotional state and generate empathetic dialogue with validation, normalization, and reframing.

Gap: These systems detect and respond to emotion in the moment, but do not store it structurally to inform future interactions.

EmoLLMs (Liu et al., 2024)¶

Paper: arXiv:2401.08508 First series of open-source LLMs for comprehensive affective analysis. Includes AAID (234K samples) and AEB benchmark (14 tasks). Outperforms ChatGPT/GPT-4 on most tasks.

EmotionPrompt & NegativePrompt¶

EmotionPrompt (Li et al., 2023, arXiv:2307.11760): Emotional stimuli in prompts produce +8% on Instruction Induction, +115% on BIG-Bench.

NegativePrompt (Wang et al., 2024, arXiv:2405.02814): Negative emotional stimuli: +12.89% and +46.25% respectively.

Implication: Emotional context significantly modifies LLM behavior. A library that maintains and injects emotional context into system memory can systematically improve response quality.

Latent Emotional Representations in LLMs (Zhang & Zhong, 2025)¶

Paper: "Decoding Emotion in the Deep" (arXiv:2510.04064)

Key findings: - LLMs develop a well-defined internal geometry of emotion (structured latent representations) - The emotional signal peaks at mid-network layers - The initial emotional tone persists for hundreds of tokens

Revolutionary implication: LLMs already represent emotion internally in a structured way. An emotional memory library can hook into these latent representations rather than building everything from the outside.

4. Gap Analysis and Positioning¶

Gaps identified in the current landscape¶

Gap	Impact	Our response
Emotion as a separate sector, not a pervasive dimension	Every system misses the affective impact on encoding/retrieval	Pervasive emotional tagging on every memory
Static representation (point in space)	Does not capture dynamics — where you are going	Layer 2: Affective Momentum (derivatives)
No global mood field	Current mood does not weight retrieval in a structured way	Layer 3: MoodField
No integrated appraisal	Emotion is assumed as input, not generated	Layer 4: Appraisal Vector
Single-signal retrieval	Semantics only, or semantics + one static emotional signal	Multi-signal retrieval with 6 components and adaptive weights
No reconsolidation	Retrieval cannot update the emotional tag	Post-retrieval reconsolidation window
Uniform decay	Emotionally intense memories are not systematically privileged	Non-linear decay modulated by arousal

Our positioning¶

                    Breadth of emotional coverage
                    (how many aspects of emotion are covered)
                              Low        High
                         ┌─────────────────────┐
                    High  │              ★ Our  │
Depth of                 │              library │
integration              │  Memory Bear AI      │
(how pervasive emotion   │  OpenMemory          │
is within the system)    │                      │
                    Low   │ MemGPT, Mem0, Zep   │
                         └─────────────────────┘

In the landscape analyzed here, AFT occupies a relatively rare position: high depth of affective integration (emotion weights both encoding and retrieval) and multi-layer coverage (core affect + momentum + mood field + appraisal + resonance). This map is interpretive, not proof of exclusivity: systems such as Memory Bear AI and OpenMemory cover important subsets of the same design space with different abstractions.

Bibliographic Notes¶

Packer, C. et al. (2023). "MemGPT: Towards LLMs as Operating Systems." arXiv:2310.08560.
Mem0 paper. arXiv:2504.19413.
Park, J.S. et al. (2023). "Generative Agents: Interactive Simulacra of Human Behavior." arXiv:2304.03442.
OpenMemory / Cavira: https://github.com/CaviraOSS/OpenMemory
Memory Bear AI: arXiv:2512.20651, arXiv:2603.22306.
Huang et al. (2024). "Emotional RAG." arXiv:2410.23041.
Xu, W. et al. (2025). "A-MEM: Agentic Memory for LLM Agents." NeurIPS 2025. arXiv:2502.12110.
MemEmo (2026). arXiv:2602.23944.
Liu, Z. et al. (2024). "EmoLLMs: A Series of Emotional Large Language Models." arXiv:2401.08508.
Li, C. et al. (2023). "EmotionPrompt: Elevating LLM Performance Through Emotional Intelligence." arXiv:2307.11760.
Wang et al. (2024). "NegativePrompt." arXiv:2405.02814.
Zhang & Zhong (2025). "Decoding Emotion in the Deep." arXiv:2510.04064.
Dynamic Affective Memory Management (2025). arXiv:2510.27418.