Long-Term Personal Memory in AI

• Long-Term Personal Memory (LPM) is a structured system in AI that persistently encodes user-specific knowledge, experiences, habits, and routines.
• It integrates semantic, episodic, habitual, and procedural memories to enable adaptive reasoning and continuity across extended, multi-modal interactions.
• LPM research, exemplified by LifeBench, showcases advanced simulation methodologies and evaluation benchmarks to drive personalized, context-aware AI agents.

Long-Term Personal Memory (LPM) is the persistent, structured, and evolving memory store within AI agents that encodes, accumulates, and manages user-specific knowledge, experiences, habits, and procedural routines across extended timescales and multiple interaction modalities. LPM systems are foundational to personalized assistants and autonomous agents, enabling robust continuity, contextualization, and adaptive behavior in long-horizon, multi-source interactions (Cheng et al., 4 Mar 2026). This entry synthesizes the state of LPM as defined and evaluated by the LifeBench benchmark and related research, detailing its memory taxonomy, generative methodologies, task suite, technical architectures, evaluation paradigms, and implications for future personalized AI.

1. Taxonomy of Memory Systems in LPM

LPM formalizes and operationalizes four distinct classes of human-inspired memory types to support life-long learning and contextual reasoning in AI agents:

• Semantic Memory: Encodes factual knowledge about the world or the persona (e.g., name, occupation, favorite foods). In LifeBench, it defines the persistent schema underlying each agent’s profile and core demographic fields.
• Episodic Memory: Captures time-stamped, event-based experiences (e.g., "On May 5 I had a work dinner in Wan Chai"), represented as hierarchical event logs mapped onto calendar time by simulated daily activities and major events.
• Habitual (Non-Declarative) Memory: Models learned, often unconscious routines such as recurring morning runs or weekly social gatherings. LifeBench injects these using survey-derived habit sampling and probabilistic, recurring atomic events.
• Procedural Memory: Embodies operationalized task knowledge (e.g., paying bills, multi-step activity planning). Procedural traces in LifeBench are implicit in multi-step event chains, decomposed by the agent’s planner (e.g., inspection → reservation → tasting) (Cheng et al., 4 Mar 2026).

This taxonomy mirrors advances in cognitive systems and is central to constructing memory-rich, agentic simulations that span semantic, episodic, habitual, and procedural dimensions.

2. Data Generation and Simulation Methodologies

Scalable and high-fidelity LPM simulations depend on efficient event synthesis, grounded real-world priors, and rigorous quality controls:

• Partonomic Hierarchy: Events are structured in a tree (partonomy) rooted in ~50 high-level "plot outlines" (work, health, finance). Each is expanded into thematic (monthly), sub-event, and ultimately atomic (<1 day) events by recursive LLM decomposition. Parallelized depth-first simulation accelerates planning (from ~2 hours to ~30 minutes for a year) and ensures daily-level context aligns with high-level intentions.
• Persona and Social Priors: Agent demographics and lifestyle attributes are sampled from anonymized surveys; social networks (20–30 members) are LLM-generated but strictly constrained by factual and relational lookup tables.
• Geographic and Calendar Fidelity: All locations and venues are validated via external map APIs; holiday calendars realistically modulate event density.
• Dual-Agent Simulation: A Subjective Agent proposes activities using a three-tiered memory model (long-term, short-term, episodic), while an Objective Agent validates their logistical and temporal feasibility using API checks for travel, availability, and time conflicts.
• Multimodal Artifacts: Synthetic phone records (calls, messages, photos, notes, health logs) and noise events are interleaved to model real digital sparsity and artifact diversity.
• Quality Assurance: Automated metrics confirm semantic, relational, geographic, and temporal authenticity. Manual rubric scoring ensures plausibility and event diversity (Cheng et al., 4 Mar 2026).

3. QA Task Suite and Memory Reasoning Benchmarks

LifeBench defines a comprehensive QA taxonomy spanning both declarative and non-declarative domains, with each question type targeting a specific mode of LPM:

• Information Extraction (IE): Single-hop factual queries ("When was my gym session on June 3?").
• Multi-hop Reasoning (MR): Aggregations across memory types and sources ("How many team meetings in Q1?").
• Temporal & Knowledge Updating (TKU): Queries about evolving facts or states ("How many swims since April?").
• Non-Declarative Reasoning (ND): Assessment of habits and routines ("What is my usual Saturday morning routine?").
• Unanswerable (UA): Abstention tasks where information is missing ("How many bottles of milk did I drink on May 8?").

Tasks are provided as open-ended short answers or multiple-choice with abstention. Evaluation is performed both overall and by category, explicitly reporting retrieval and reasoning accuracy per LPM facet (Cheng et al., 4 Mar 2026).

4. Evaluation Metrics and System Performance

Multiple accuracy and data-quality metrics are used to map LPM agent performance and simulation fidelity:

• Overall and Per-Category Accuracy: Accuracy = (Number of correct answers) / N. Detailed breakdown by IE, MR, TKU, ND, UA.
• Relation Consistency: Acc_person = 100 × (number of consistent relation mentions) / (total relation mentions).
• Location and Trip Authenticity: Geospatial correctness measured by verified location authenticity and trip congruence within strict temporal/spatial tolerances.
• Event Diversity: Normalized Shannon entropy (H_norm), Simpson Diversity Index, quantifying event heterogeneity.
• System Performance: On LifeBench, SOTA systems achieve 55.2% overall accuracy (MemOS), with lower scores for multi-hop, temporal, and non-declarative queries. Category-level accuracy for MemOS: IE ≈ 75%, MR ≈ 60%, TKU ≈ 50%, ND ≈ 40%, UA ≈ 30%. Key failure modes include retrieval errors (temporal misalignment), incomplete evidence aggregation, hallucinated details, temporal/causal confusion, and omission of salient but "minor" facts (Cheng et al., 4 Mar 2026).

5. Technical and Architectural Implications for Agents

Agent architectures targeting LPM capability must integrate several technical strategies:

• Joint Modeling of All Memory Types: Architectures must support seamless integration of semantic, episodic, habitual, and procedural memory and switch contextually among them during retrieval and reasoning.
• Temporal Indexing and Time-Aware Retrieval: Temporal information must be encoded and leveraged beyond naive similarity search to correctly constrain before/after and causal dependencies.
• Multi-Source Data Fusion: Structured data from phone artifacts (e.g., GPS traces, health logs, message metadata) must be aligned or joined with text-centric memories for robust, multi-modal reasoning.
• Proactive Memory Management: Given the long horizon and data volume, agents must make explicit decisions about which memories to store, summarize, or prune to maintain tractable and high-utility LPM substrates.
• Evaluation-Driven Development: Rigorously annotated benchmarks guide algorithmic improvements and expose key errors in consolidation, retrieval, and long-horizon reasoning (Cheng et al., 4 Mar 2026).

6. Limitations and Future Directions

Despite progress, LifeBench and comparable LPM benchmarks expose structural challenges:

• Temporal Reasoning and Evidence Aggregation: Multi-hop and long-span questions depress accuracy, indicating persistent limitations in consolidating and updating across temporally dispersed events.
• Non-Declarative Memory: Modeling and querying routines and procedural knowledge remain lower-performing and less well-defined than semantic/episodic memory.
• Sparsity and Noise: Realistic simulation must further address digital trace sparsity and event noise to match real-world memory artifacts.
• Scalability: Parallel simulation and efficient event decomposition are required for scaling agent traces to lifespans spanning multiple years or modalities.
• Improved Proactive Forgetting and Summarization: Agents should better model dynamic forgetting, salience-sensitive pruning, and continual abstraction, especially as memory stores expand (Cheng et al., 4 Mar 2026).

LPM research, exemplified in LifeBench, is thus driving the convergence of cognitive-inspired multi-system memory modeling, realistic behavioral simulation, and systematic evaluation, enabling the future of adaptive, contextually aware, and personalized AI agents.