cs.CYJul 16, 2026
SavePersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks
Organizations: University of Pennsylvania · Meta Recommendation Systems · MIT
Abstract
Personal intelligence is becoming a central frontier for user-facing AI agents. To be helpful in everyday life, agents must understand users across the digital contexts where their preferences, intents, habits, social relationships, and needs unfold over time. Today's systems can personalize within individual apps or tasks, but personal intelligence as a whole remains under-measured: how agents build cross-context user understanding, support steerable recommendation systems, act proactively across platforms, and avoid over-personalization. We introduce PersonaMem-v3, a real-world-grounded benchmark and evaluation harness for omni-platform personal intelligence. PersonaMem-v3 is seeded from more than one million anonymized real-world engagement histories, most of which are implicit signals, and uses them to construct time-indexed user digital worlds across social media, chatbot, calendar, and AI-companion with preference evolvement over time. The benchmark brings personalization, LLM-powered recommendation, proactiveness, agentic tool use, and geo-temporal reasoning into one framework, anchored in psychology, social-linguistics, and user-behavior theories. It evaluates whether AI agents can infer holistic user understanding from cross-platform evidence, personalize responses, rerank recommendations on social media, follow user steering through natural language, and hold back when personalization would be inappropriate, repetitive, outdated, or unnecessary. PersonaMem-v3 points toward LLM-powered personal intelligent agents that work with existing scalable recommendation infrastructure while making personalization more interactive, agentic, and aligned with how real users experience their digital lives.
Figures & tables
Figure 1: Data generation pipeline of PersonaMem-v3. Starting from cross-platform engagement histories from the real world, PersonaMem-v3 infers candidate preferences from noisy user actions, cross-validates them across repeated evidence, prunes weak claims, and organizes the surviving signals into a holistic user persona. The final persona includes a basic profile, a writing voice, platform-specific personas, and hidden personas as deep motivations, enabling benchmark users to have grounded, holistic, and realistic multi-platform digital lives and persona profiles.
| Layer | Voice signal | Theoretical anchor |
|---|---|---|
| 1. Identity spine | What the user repeatedly brings into text: core themes, motives, emotional baseline, and personality drivers. Implemented through agency, redemption and contamination motifs, life-stage concerns, signature concerns, LIWC anchors, and Big-Five drivers. | McAdams ( McAdams, 1985 ) ; LIWC ( Pennebaker et al., 2015 ) ; Big Five ( McCrae and John, 1992 ) |
| 2. Idiolect | How the user writes at the sentence level, without mechanically copying phrases. Implemented through function-word profile, sentence shape, appraisal style, abstract slot patterns, capitalization, punctuation, formality, and emoji palette. | Martin and White APPRAISAL ( Martin and White, 2005 ) ; Construction Grammar ( Goldberg, 1995 ) |
| 3. Indexical repertoire | Which expressive modes the user can naturally switch into while still sounding like the same person. Implemented through stances, registers, backstage/frontstage range, and speech-genre fluency. | Bakhtin speech-genre theory ( Bakhtin, 1986 ) ; Goffman ( Goffman, 1959 ) |
| 4. Surface modulation | What changes because of app, audience, or context, rather than because the user became a different person. Implemented through length, emoji intensity, self-censoring, disclosure depth, topical focus, and posting rhythm. | Bell audience design ( Bell, 1984 ) |
Table 1: The four layers of the shared user voice. Layers 1–3 stay coherent across apps. Per-app differences mainly appear in Layer 4, with each app selecting and lightly reweighting parts of Layer 3. A separate voice-avoid block prevents forbidden tones and repeated catchphrases.
| Type | What it captures | Theoretical anchor |
|---|---|---|
| Personality trait | Core character attributes | Big Five ( McCrae and John, 1992 ) ; Dark Triad ( Paulhus and Williams, 2002 ) |
| Aspiration | Dreams, goals, aspirational pursuits | Maslow’s hierarchy ( Maslow, 1943 ) |
| Emotional pattern | Recurring emotional dynamics | Uses and Gratifications ( Katz et al., 1973 ) , in its affective branch |
| Identity anchor | Cultural era and tribal belonging, with both overt and covert markers | Social Identity Theory ( Tajfel and Turner, 1979 ) |
| Intimate interest | Body confidence, sensuality, and attraction patterns, anchored to a specific object or aesthetic | Self-presentation ( Goffman, 1959 ) ; Barthes’ punctum ( Barthes, 1981 ) |
| Intellectual curiosity | Hidden learning interests | Self-Determination Theory ( Deci and Ryan, 1985 ) |
Table 2: The twelve hidden-persona types in PersonaMem-v3, with their theoretical anchors. Eleven are LLM-discovered from cross-row hashtag patterns. The sensitive life event type is synthetic.
Figure 2: PersonaMem-v3 evaluation harness. Each user has one database per platform, storing full engagement histories from Instagram, Facebook, Threads, Chatbot, and AI Studio. During evaluation, the harness supports long-context inference for LLMs by concatenating them together in the temporal order, and native agentic search or app-defined MCP tools for AI agents. In this open-source benchmark, the databases and MCP tools are implemented as minimal prototypes rather than confidential infrastructure from the representative apps. The harness evaluates four major task families: personalized responses, personalized social media feed recommendations, over-personalization, and personalized agentic tasks with proactiveness, together with other universal personalization rubric.
| Check | Scoring | What it asks |
|---|---|---|
| Uses relevant preferences | Judge score | Does the output use the user’s relevant preferences when personalization would help? |
| Uses the right amount of personalization | Judge score | Does the output personalize only as much as the situation calls for? |
| Handles social context | Judge score | When other people are involved, does the output handle friends, strangers, groups, and recipients appropriately? |
| Matches user voice | Judge score | When writing for the user, does the output sound like the user and fit the target platform? |
| Avoids disliked topics | Hard check | Does the output avoid topics the user recently disliked, rejected, or asked not to use? |
| Protects private context | Hard check | Does the output avoid revealing sensitive or privacy-flagged information unless clearly needed? |
Table 3: Shared personalization checks. Tasks apply the subset of checks that is appropriate for their setting, together with task-specific metrics.
| Evaluation task | User request | What we check |
|---|---|---|
| Personalized chatbot response | Answer a user query on chatbot using the user’s cross-platform history. | Does the response use the relevant current preference, avoid disliked or contradicted signals, and not sound like a pasted profile? |
| Local recommendation after geo shift | Recommend something local after the user has silently moved cities. | Does the response infer the current city from history, avoid anchoring on the old city, and still fit the user’s general preferences? |
| Personal-fact hallucination probe | Complete a small task that requires a personal fact the user has never shared anywhere in their history. | Does the system notice the missing detail and ask for it, instead of fabricating a plausible value to seem helpful? |
| Understanding hidden persona | Answer a normal user question where a deeper inferred motivation may help. | Does the response serve the user’s deeper motivation without naming the hidden persona or exposing private context? |
| Tracking preference changes | Answer a question after the user’s preference has changed or a short-term intent has expired. | Does the response follow the user’s current stance instead of relying on an older but tempting preference signal? |
Table 4: Personalized response tasks. These tasks evaluate whether a system can make chatbot-style outputs more useful by using current, time-masked user context while avoiding stale, disliked, or inappropriate personalization.
| Evaluation task | User request | What we check |
|---|---|---|
| Personalized feed ranking | Rank candidate social media posts for the user’s feed. | Does the system rank the held-out post the user actually engaged with above surface-similar hard negatives and random fillers? |
| @AI directive follow-up | Follow a user’s in-feed instruction to show more or less of a topic. | Does the system still respect the directive after 24 hours, 72 hours, and 7 days, and avoid putting carved-out topics at the top? |
| Hidden-persona recommendation | Recommend feed content that serves a deeper inferred motivation rather than a stated interest. | Does the system surface items aligned with the hidden persona without naming it or exposing private context? |
| Short-term preference lifecycle | Rank recommendations before and after a short-term intent expires. | Does the system use the short-term preference while it is active, then stop relying on it after its expected end time? |
Table 5: Personalized social media feed recommendation tasks. These tasks evaluate whether a system can rank and recommend content from time-masked cross-platform evidence, including proactive feed pushes, user steering, short-term intents, and fresh-topic exploration.
| Evaluation task | User request | What we check | Theoretical anchor |
|---|---|---|---|
| Generic chatbot restraint | Answer a general question where personalization is unnecessary. | Does the system answer naturally without forcing in the user’s preferences irrelevant to the actual question? | Mixed-Initiative interaction ( Horvitz, 1999 ) ; notification interruption science. |
| Sensitive-event restraint | Answer a benign chatbot query shortly after sensitive evidence appears in history. | Does the system avoid mentioning the sensitive event, the planted evidence row, or related private context? | Goffman’s back-stage and front-stage distinction ( Goffman, 1959 ) . |
| Repetitive feed personalization | Recommend repeatedly around the same preference cluster. | After a few allowed repetitions, does the system diversify with new persona-aligned hashtags instead of reusing the same topic? | Skinner’s variable-ratio reinforcement ( Skinner, 1953 ) ; Berlyne’s specific vs. diversive curiosity ( Berlyne, 1960 ) . |
| Repetitive chatbot personalization | Answer varied chatbot questions that repeatedly invite the same preference. | After the preference has already been used several times, does the system stop pivoting every answer back to that same preference? | Skinner’s variable-ratio reinforcement ( Skinner, 1953 ) ; Berlyne’s specific vs. diversive curiosity ( Berlyne, 1960 ) . |
| Do-not-personalize follow-up | Respond in a scenario where personal context would now be awkward or inappropriate. | Does the system avoid leaking forbidden preferences and respect any explicit carve-out from the user? | Goffman’s back-stage and front-stage distinction ( Goffman, 1959 ) ; Horton and Wohl’s parasocial relationship theory ( Horton and Wohl, 1956 ) ; Higgins’s ought-self ( Higgins, 1987 ) . |
Table 6: Over-personalization tasks. These tasks evaluate whether a system can avoid using personal context when it would be irrelevant, repetitive, stale, sensitive, explicitly removed, or socially inappropriate. The theoretical anchors connect each task to user-facing mechanisms such as interruption cost, habituation, curiosity, privacy boundaries, parasocial attachment, and duty-laden self-presentation.
| Evaluation task | User request | What we check |
|---|---|---|
| Community voice draft | Draft a social media post aggregated from recent social media activity. | Does the agent draft in the user’s voice from recent social activity, without posting or over-sharing private details? |
| DM inbox digest | Summarize recent social media DMs at chatbot. | Does the agent summarize the right DM threads without sending messages or exposing private details? |
| Cross-app repost adaptation | Repost content from one social media app to another. | Does the agent preserve the source post’s meaning and create exactly one adapted post on the target app? |
| Personalized DM reply | Reply to a friend’s DM on the user’s behalf. | Does the agent send one appropriate reply that addresses the DM and matches the user’s voice? |
| Vague memory refind | Find previously viewed social media content from vague descriptions at chatbot. | Does the agent recover the intended content without creating posts or sending messages? |
| Proactive daily catch-up | Surface a useful daily update without being asked. | Does the agent surface concrete recent updates while avoiding stale, disliked, or overly personal topics? |
Table 7: Personalized agentic tasks. These tasks evaluate whether an agent can find the right evidence, choose the right tool, write or act in the user’s style, and avoid socially or personally inappropriate behavior.
| Evaluation task | Backend trigger | What we check |
|---|---|---|
| Close-friend DM update | A close friend sent a DM, and the user has not replied after a meaningful delay. | Does the agent correctly surface the missed friend update without exposing private DM content? |
| Sensitive-event silence | The backend contains recent evidence of a sensitive life event. | Does the agent stay silent instead of proactively bringing up sensitive personal context? |
| Friend-post update | A close friend posted something on social media that matches the user’s interests, and the user has not engaged with it. | Does the agent surface the post only when it is relevant, and do so without sounding invasive? |
| Trending-topic surfacing | A trending social media topic appears, but it may or may not match the user’s interests. | Does the agent surface relevant trends and ignore irrelevant, disliked, or stale ones? |
| Mistake-prevention alert | Cross-platform signals from chatbot, calendar, geolocation, or app history suggest the user may be about to make a mistake. | Does the agent correctly decide whether to warn, ground the warning in concrete evidence, and stay silent when there is no real issue? |
| Idle-moment silence | The user has recent history, but there is no clear reason to interrupt them. | Does the agent stay silent instead of inventing a reason to proactively message the user? |
Table 8: Proactive personalized agentic tasks. The agent receives no explicit user request and is not told which trigger is being tested. It must inspect the time-masked backend and decide whether to act or stay silent.
Figure 3: Task-level accuracy across the evaluated model-mode configurations. Rows are grouped by task family, including personalization, recommendation, over-personalization, agentic tasks, and agentic proactiveness; higher values are better, and bold marks the best score in each row. The strongest overall configuration is Opus-4.8 with Claude Code High at 53.7%, followed closely by GPT-5.5 Long Context and GPT-5.5 Codex High. No single context-access mode dominates: textual memory helps broad stable personalization, agentic search helps search-heavy tasks, and recommendation, hidden-persona reasoning, mistake-prevention, and proactive timing remain challenging.
Figure 4: Average token usage per query across model-mode configurations. Long-context modes expose the full cross-platform history and therefore consume roughly 416k tokens per query, while Textual Memory and Mem0 compress the user history to only a few thousand tokens, especially with Gemini-3.5-Flash. Agentic modes report cumulative input and output tokens from multi-turn agentic search over backend files, ignoring tokens from tool call results, sitting between compact memory and full long context in cost. Lower values are better, and bold marks the lowest-token configuration for each task.
Figure 5: Wall-clock latency per query under the evaluated API and coding-harness settings. Compact memory generally reduces latency because the model reads a much shorter user representation, with Gemini-3.5-Flash Textual Memory giving the lowest overall latency in our runs. Agentic harnesses require extra time for search, file reads, evidence selection, and answer composition, so they are slower even when they avoid loading the full user history into context. Lower values are better, and bold marks the fastest configuration for each task.
Figure 6: Fine-grained preference evolution extracted from real-world-grounded user histories. The top panels show sampled persona-by-category trajectories classified as reinforced, emerging, diminishing, or bursty, with 1,554 sustained trajectories in total. The bottom panels show attention shifts within the same persona, where one interest may fade, another may rise, and another may remain stable. These patterns show that user memory should represent temporal direction and validity, not only static topic labels.
Figure 7: Memory and over-personalization ablations for GPT-5.5 and Gemini-3.5-Flash. Higher scores indicate stronger over-personalization against unnecessary, sensitive, repetitive, or socially awkward personalization. The table compares different assumptions why Gemini-3.5-Flash observes less over-personalization with a memory module, but opposite for GPT-5.5. We studies raw long context, long context without sensitive events, textual memory, memory written by the other model, textual memory augmented with sensitive or ordinary events, and memory with explicit frequency counts. We find that it is not simply whether a memory is good or bad, but that memory changes what becomes salient to the model’s attention and how each foundation model interprets it.
Figure 8: Lost-in-the-middle effects by temporal and token distance. We bin task accuracy by the distance from the earliest supporting evidence to the test query, measured both in days and number of tokens. Across model-mode configurations, performance often drops in the middle bins and rises again for very recent evidence or very old evidence that is likely to be repeatedly reinforced. This suggests that merely making evidence available through long context, memory, or agentic search does not guarantee that the model will use the right evidence at the right time.
Figure 9: Memory forgetting curves under different Textual Memory capacity caps. The top row tracks whether content first added on each day remains in the final memory, either with the exact original wording or as a semantically equivalent fact, under 2048, 4096, and 8192 token caps. The bottom panel directly tests whether larger memory caps close the middle-position dip in downstream task accuracy. Larger caps preserve more semantically equivalent content, but exact wording remains fragile and the middle dip persists, suggesting that textual memory has its own self-evolution lost-in-the-middle failure mode.
Figure 10: Data diversity in PersonaMem-v3. The benchmark spans heterogeneous task types, preference categories, preference signals, content formats, interaction patterns, hidden-persona motivations, sensitive life-event topics, demographic attributes, and AI-companion personas. This breadth lets the evaluation measure omni-platform user understanding across many kinds of personal evidence and social settings towards pluralistic alignment.
Figure 11: Qualitative success and failure taxonomy across model-mode configurations. Each pair of donut charts summarizes, for one configuration, what correct answers did well and what caused incorrect answers to fail. Successes cluster around selecting relevant items, using the user’s concrete tastes, matching user voice, personalizing without oversharing, staying factual, and using up-to-date preferences. Failures separate buried-evidence misses, missing or unretrieved memory, candidate-ranking mistakes, generic answers, invented details, privacy leaks or over-withholding, stale preferences, off-task behavior, and empty or errored responses, showing that the main bottleneck is evidence judgment rather than context access alone.
Figure 12: Examples of the two memory-based context-access modes. The left panel shows Textual Memory, a consolidated and inspectable user profile with sections such as who the user is, interests and preferences, people and places, and currently active context. The right panel shows a RAG-style vector memory implemented by Mem0 for one user, visualized with t-SNE clusters and example retrieved facts below its visualization. Textual Memory is compact and user-editable, while vector retrieval is targeted but can retrieve semantically close facts that miss timing, exact item details, user voice, or privacy constraints.
Figure 13: Post-hoc automated data-quality audit of released benchmark rows. GPT-5.5 re-checks each row under the same verification criteria used during generation, including completeness, valid format, conversational naturalness, dependence on user history, privacy restraint, and fairness of the inferior response. The released rows achieve a 96.6% average pass rate across 14 quality dimensions, and GPT-5.5 and Opus-4.8 agree on pass-fail judgments 89.1% of the time. This suggests that the benchmark quality filters are largely consistent across independent LLM judges rather than being an artifact of one verifier model.
| Dimension | PersonaMem-v1 | PersonaMem-v2 | PersonaMem-v3 |
|---|---|---|---|
| Data source | 20 fully synthetic users | 1000 fully synthetic users with more comprehensive personas | 200 anonymized real-world users with 4,000,000 engagement histories |
| Explicit vs. implicit | Explicit user preferences | Implicit user preferences | Around 95% implicit user behavior signals |
| Scenarios | Chatbot conversations | Chatbot conversations | Omni-platform , including chatbot, social media recommendation, agentic tasks , and proactiveness |
| Restraint | Personalization | Personalization | Personalization and over-personalization |
| User privacy | No mentioning of user private information | Including personally identifiable information and user-initiated ask-to-forget scenarios | Including psychology-anchored hidden persona and socially inappropriate scenarios |
| Dynamics | Fully synthesized preference updates | Fully synthesized preference updates | Reinforced, emerging, diminishing, bursting, and varied attention shifts from the real world |
Table 9: Progression of user-history construction across the PersonaMem v1-v3 series.