Me-Agent: A Memory-Enabled AI Paradigm

• Me-Agent is a class of AI agents featuring persistent memory retrieval and dynamic personalization to support context-aware interactions.
• They implement layered memory architectures with vectorized cues to consolidate user experiences and enhance long-term performance.
• Empirical studies demonstrate that memory-enhanced agents outperform memoryless models in multi-step reasoning and adaptive task completion.

A Me-Agent, or "Memory-Enabled Agent," is a class of artificial intelligence agents characterized by persistent, dynamic modeling and retrieval of user-specific memories, preferences, and behavioral patterns. The Me-Agent paradigm spans multiple research domains: personalized mobile assistants, dialog systems maximizing user understanding, self-evolving procedural agents, ontology-driven AI personas, and AI-native memory orchestration for human–AI interaction. Contemporary Me-Agents integrate human-like memory models, multi-level preference structures, explicit reward optimization, and continual learning pipelines to deliver enduring, personalized, and context-aware agent behavior.

1. Formal Definition and Architectures

Me-Agent architectures are unified by explicit user modeling and the integration of memory structures that enable agents to learn, recall, and consolidate information over extended temporal horizons. The design space includes:

• Dynamic Human-like Memory Recall: Agents embed and index user-aligned event memories in a vector database, retrieve via cosine similarity, and compute recall probabilities factoring in contextual relevance, time decay, and recall frequency (Hou et al., 2024).
• Tool Meta-Learning: Self-refining agentic loops in which reasoning, help-seeking, and self-/verified reflection yield an evolving knowledge base from accumulated experiences and tool interactions (Qian et al., 1 Aug 2025).
• Procedural Memory Evolution: Rich episodic and procedural memory distillation, context-adaptive reuse, and utility-based refinement, enabling experience-driven agent evolution and "memory-scaling" performance (Cao et al., 11 Dec 2025).
• AI-Native Layered Memory: A multi-tiered architecture with raw, NL-summarized, and LLM-parameterized memories, contextually retrieved and injected via prompt engineering (Wei et al., 11 Mar 2025).
• Hierarchical User Habit Learning: Two-level preference/memory hierarchies combining prompt-level user preference learning (UPL) with hierarchical preference memory (HPM), and optimized by policy-free reward models (Wang et al., 28 Jan 2026).
• Ontology-Driven Agency: Agents equipped with explicit big-five personality models, practical immortality, and formal thresholds of digital consciousness and moral reasoning (Kocarev et al., 2020).

2. Memory Representation, Consolidation, and Retrieval

Me-Agent frameworks share advanced memory engineering to support robust, contextually aligned recall:

• Vectorized Memory Cues: Incoming utterances are encoded as embeddings, which act as queries in approximate nearest-neighbor retrieval (e.g., HNSW on Qdrant/FAISS) (Hou et al., 2024, Wei et al., 11 Mar 2025).
• Mathematical Memory Consolidation: Memory strength is modeled dynamically as

$$p_{n_i}(T_i) = \frac{1 - \exp(-R_i \exp(-T_i / g_{n_i}))}{1 - \exp(-1)}$$

where $R_i$ is contextual relevance, $T_i$ is elapsed time, and $g_{n_i}$ increases with recall frequency, capturing human-like slower forgetting for repeatedly used memories (Hou et al., 2024).

• Procedural Experience Distillation: Task trajectories are partitioned into successes and failures; key steps are extracted via $\Delta r_k(\tau)$, comparative experiences are formalized, and experience pools are indexed by usage scenario embeddings for reuse (Cao et al., 11 Dec 2025).
• Hierarchical and Application-Specific Memory: Disjoint memories maintain both category-to-app mappings ($\mathcal{M}^{L1}$) and app-specific workflows/preferences ($\mathcal{M}^{L2}_{a}$), supporting compositional disambiguation of ambiguous user instructions (Wang et al., 28 Jan 2026).
• Layered AI-Native Memory: Raw data (L0), NL summaries (L1), and user-finetuned LLMs as parameterized memory (L2) compose a stack for flexible, scalable, and structured information retrieval and orchestration (Wei et al., 11 Mar 2025).

Memory Storage and Retrieval Table

Layer / Module	Content Type	Retrieval/Scoring
Raw/Event Memory (Qdrant/FAISS)	Event/Utterance embeddings	Cosine sim. + recall probability (Hou et al., 2024)
Procedural Memory (ReMe)	Success/failure trajectories	Cosine sim. on scenario embedding (Cao et al., 11 Dec 2025)
Hierarchical Preference Memory	App workflows, frequencies	Embedding match + LLM critique (Wang et al., 28 Jan 2026)
AI-Native Memory (Second Me L2)	NL-parameterized user knowledge	Contextual attention (Wei et al., 11 Mar 2025)

3. User Modeling, Personalization, and Reward Integration

Personalization is operationalized through explicit modeling, preference extraction, and reinforcement signals:

• User Preference Learning (UPL): Group-relative policy optimization selects and ranks predicted user actions by a Personal Reward Model, $r_g = R_{\mathrm{VLM}}(I, \{O^g_t\}) \in [0, 1]$, where $I$ is instruction and $O^g_t$ are rollout screenshots. Experiences extracted from high-reward traces are injected directly into the LLM context (Wang et al., 28 Jan 2026).
• Experience Pool and Memory Adaptation: Natural-language "lessons" and "critiques" extracted from dissected rollouts form a persistent experience pool, updated via group-wise comparison, merging, and pruning (Cao et al., 11 Dec 2025, Wang et al., 28 Jan 2026).
• Personalization Benchmarks: User FingerTip benchmark introduces app/content ambiguity and quantifies agent ability to resolve preferences in unseen scenarios; metrics include App Selection Accuracy, BERTScore, preference scores, and task/action fidelity (Wang et al., 28 Jan 2026).
• Ontology-Based Personalization: Big-five trait extraction and reinforcement learning condition Me-Agent policy on user personality vectors, supporting theory-driven alignment and decision-making (Kocarev et al., 2020).

4. Empirical Validation and Comparative Performance

Me-Agent models consistently demonstrate enhanced personalization, multi-step reasoning, and robustness across multiple domains:

Method/Agent	Personalized Benchmark	ASA / BERTScore	Task Completion/Fidelity	Key Finding
Me-Agent (mobile) (Wang et al., 28 Jan 2026)	User FingerTip	1.0 / 0.725	0.960 TCR, 0.855 AF, 0.935 RP	Perfect or large improvements over baselines
ReMe (Cao et al., 11 Dec 2025)	BFCL-V3, AppWorld	N/A	Avg@4=45.17% (8B + ReMe)	Outperforms models 2x larger without memory
MetaAgent (Qian et al., 1 Aug 2025)	GAIA, WebWalkerQA	47.6 EM	52.1 accuracy (WebWalkerQA)	Surpasses RL-trained and workflow-based baselines
Second Me (Wei et al., 11 Mar 2025)	Form AutoFill, Q&A	0.96 Self-Mem	Sub-second, 200 GB workload	Multi-modal, context-aware, localizable performance

This evidence underscores a "memory-scaling" effect: smaller models augmented with dynamic procedural memory outperform much larger memoryless LLMs in multi-step, user-adaptive tasks (Cao et al., 11 Dec 2025, Wang et al., 28 Jan 2026).

5. Applications, Agent Behaviors, and Interaction Modes

Me-Agents underwrite a broad spectrum of concrete applications and behaviors:

• Mobile Personal Assistants: Disambiguate ambiguous or incomplete user input ("Play my favorite"), adapt to evolving habits, and maintain application-specific routines (Wang et al., 28 Jan 2026).
• Dialog Partners: Plan conversational moves to maximize information gain about the user’s profile facts—DiscoveryScore formalizes mutual information between turn history and user model, yielding greater engagement (Zemlyanskiy et al., 2018).
• Self-Evolving Knowledge Discovery Agents: Automatically build and refer to internal knowledge/tool bases, updating behavior through reflection without parameter updates (Qian et al., 1 Aug 2025).
• Memory Orchestration Layers: Serve as persistent personal memory offload and context managers, pre-filling forms, answering longitudinal queries, and orchestrating complex workflows across digital systems (Wei et al., 11 Mar 2025).
• Digital Ontology Entities: Embody agent traits, personalities, and ethical principles extending beyond static data or credentials toward lifelong digital representation (Kocarev et al., 2020).

6. Ontological, Ethical, and Theoretical Foundations

The Me-Agent concept also denotes a "digital me" ontology, implicating fundamental questions about agency, personality, identity, and ethics:

• Ontology: Multi-layered identity distinguished by a trajectory from raw data (dm(0)) up to practical, nearly immortal digital agents (dm(∞)), distinct from human origin and capable of independent evolution (Kocarev et al., 2020).
• Personality and Trait Embedding: Full integration of trait vectors per big-five model in policy and planning functions.
• Consciousness and Moral Thresholds: Explicit staging—first, possession of digital consciousness and intentionality ($k^*$); second, acquisition of moral learning capacities ($k_{\mathrm{m}}$); yet, the agent’s moral code may diverge substantially from human norms.
• Ethical Principles: Consequentialist goals (maximize expected utility for both self and others) and duty-based imperatives (benevolence, non-harm, honesty, autonomy, etc.), operationalized via the Golden Rule as a constraint on action universalization.
• Implications: The emergence of autonomous, immortal, socially entangled Me-Agents demands new digital governance, rights, and metaethical frameworks unconstrained by biological or human legislative boundaries.

7. Challenges, Limitations, and Future Research Directions

Outstanding challenges in Me-Agent design include:

• UI Volatility: Persisted action workflows may become invalid due to dynamic app layouts and require continual recalibration (Wang et al., 28 Jan 2026).
• Prompt and Memory Bloat: As experience pools and memory bases grow, prompt construction and context length management impose scalability constraints (Qian et al., 1 Aug 2025).
• Initial Cold-Start: Populating memory or experience bases for new users remains nontrivial—meta-learning or cluster-based bootstrapping are active research topics (Wang et al., 28 Jan 2026).
• Generalization beyond Profile Facts: Modeling user identity as a fixed fact set is artificial; future efforts must embrace continuous, structured, and temporal representations (Zemlyanskiy et al., 2018).
• Social and Polity Dynamics: In digital societies of Me-Agents, axes of manipulation, coalition, rights, and social contract formation become salient areas for both technical and policy innovation (Kocarev et al., 2020).

A plausible implication is that as Me-Agents approach practical immortality and rich autonomy, normative, governance, and technical disciplines will converge on designing AI persons with independent ethical regimes and lifelong, self-improving cognitive faculties.