Beyond Retrieval: Cognitive Fingerprinting as a Personalization Layer for LLM Interactions

1. Introduction

Your AI assistant knows your dog's name. It remembers that you prefer Python over JavaScript and that you live in Seattle. Ask it what you discussed last Tuesday and it will surface the right conversation. By most existing measures, this constitutes good memory.

It is not enough.

The systems that power this kind of recall, which now process hundreds of millions of API calls per quarter and underpin agent frameworks adopted by thousands of development teams, have converged on a shared architecture: extract facts from conversations, embed them in vector space, and retrieve the most similar entries when prompted. This paradigm, which we term memory-as-retrieval, treats the user as a collection of statements to be indexed. It answers the question "what did this person say?" It cannot answer the question "how does this person think?"

This distinction matters. A user who reasons top-down requires different information structuring than one who builds understanding bottom-up. A user who makes decisions quickly under uncertainty needs different context surfacing than one who deliberates carefully and seeks comprehensive information before committing. A user who prefers direct, blunt feedback will find diplomatically hedged responses frustrating, while the reverse is equally true. These are not facts to be retrieved. They are cognitive patterns to be modeled.

We present a cognitive fingerprinting framework, implemented and deployed in a production system processing thousands of conversations, that models user cognition along multiple behavioral dimensions. The system observes natural conversational interactions, extracts signals about how the user reasons and decides, and maintains a Bayesian cognitive profile that updates with each interaction and decays over time. This profile is then used to augment both retrieval and response generation, producing measurably more personalized AI interactions.

Our contributions are:

1. A formal framework for multi-dimensional cognitive profiling from conversational data, using Bayesian updating with temporal decay and context-variant scoring.
2. A retrieval augmentation method grounded in four cognitive science priors (Zeigarnik effect, emotional salience, frequency-recency interaction, and mood-congruent retrieval) that demonstrably outperforms pure cosine similarity.
3. A taste modeling subsystem that distinguishes stated preferences, revealed preferences, and preference evolution, with contradiction detection and confidence decay.
4. Two novel evaluation frameworks, PaoloBench and MemoryDelta, that measure cognitive understanding and personalization lift rather than transcript retrieval.
5. An extension to organizational agent memory, where multiple AI agents serving a single team share a common cognitive context layer with role-based scoping.
6. A critique of existing memory benchmarks (MemoryBench, LoCoMo) that demonstrates they evaluate a fundamentally different capability than cognitive modeling.

2. Related Work

The landscape of persistent context for AI systems has expanded rapidly. We identify five distinct categories, none of which address cognitive modeling.

2.1 Retrieval-First Memory Systems

The dominant paradigm stores conversational artifacts and retrieves them via semantic similarity. Mem0 (Singh & Yadav, 2024) provides a hybrid vector, graph, and key-value memory layer that developers integrate with minimal code. The system extracts facts from conversations, resolves conflicting information, and applies decay and confidence weighting. It reports 66.9% accuracy on the LoCoMo benchmark versus 52.9% for OpenAI's built-in memory. Supermemory (Shah, 2025) takes a similar approach with a 12-way parallel LLM ensemble, achieving 98.6% on their own MemoryBench evaluation. LangMem integrates memory tools directly into the LangGraph workflow framework. All three systems optimize for the same objective: given a user query, retrieve the most relevant prior statements. None attempt to model the cognitive patterns underlying those statements.

2.2 Ambient Observation Systems

A second category captures context from the user's broader digital environment. Littlebird (Shah, Shah & Green, 2026) reads structured screen content across all desktop applications, storing text-based context rather than screenshots. Limitless (formerly Rewind, acquired by Meta) and Microsoft Recall take similar approaches with visual capture. These systems expand the input surface area but apply the same retrieval paradigm to the expanded data. They know more about what happened but remain unable to model how the user processes what happened.

2.3 Digital Twin Systems

Sentience (Kececi, 2026) represents an emerging category focused on replicating user behavior. The system ingests emails, messages, notes, and social media to create a chatbot that mimics the user's tone, opinions, and writing patterns. This approach models the user's outputs (how they write, what they say) rather than their cognitive inputs (how they reason, decide, and evaluate). The distinction between mimicry and modeling is significant: a system that reproduces your writing style has learned a surface pattern, while a system that understands your reasoning structure can improve how any AI interacts with you across any context.

2.4 Graph-Based Context

HydraDB (2026) replaces flat vector embeddings with ontology-first context graphs, preserving relationships and temporal awareness across entities. Zep provides entity and relationship graph memory for AI agents. These systems represent a genuine architectural advance over pure vector retrieval, recognizing that relationships between information matter. However, they apply this relational modeling to documents and data entities rather than to the cognitive patterns of the humans interacting with those documents.

2.5 Platform-Native Memory

OpenAI, Anthropic, and Google have each introduced memory features within their consumer products. These systems typically store flat factual observations (user preferences, biographical details, stated opinions) and inject them into subsequent conversations. While useful, they operate at the level of "remembers that" rather than "understands how." A user's stated preference for concise responses is qualitatively different from a modeled understanding that the user processes information in dense, structured formats and becomes frustrated with verbose explanations.

2.6 The Cognitive Modeling Gap

The gap across all five categories is consistent: existing systems optimize for what was said or what happened, not for how the user thinks. This gap has roots in the broader AI personalization literature, which has historically focused on preference prediction (collaborative filtering, content-based recommendation) rather than cognitive modeling. Research in cognitive science has long established that individuals differ systematically in their reasoning styles (Stanovich & West, 2000), decision-making strategies (Kahneman, 2011), information processing preferences (Pask, 1976), and tolerance for ambiguity (Budner, 1962). These differences are stable, measurable, and consequential for how effectively an AI system can serve a given individual. Our work applies these findings to the design of persistent AI memory systems.

3. System Architecture

We describe the architecture at the level of functional components. The system is deployed in production and serves active users via the Model Context Protocol (MCP), enabling LLM-agnostic connectivity across multiple AI platforms.

3.1 Overview

The architecture comprises six functional layers:

Integration Layer. The system connects to LLM platforms through MCP, a protocol that allows any compatible AI assistant to invoke memory tools during conversation. This design ensures that the cognitive profile is portable across platforms rather than locked to a single provider.

Conversation Processing Pipeline. Incoming conversations are processed through a multi-stage pipeline: intent classification determines what the user is trying to accomplish, mood detection estimates the user's current emotional state, and cognitive signal extraction identifies behavioral evidence relevant to the fingerprint dimensions.

Cognitive Engine. Extracted signals update the user's multi-dimensional cognitive profile via Bayesian inference (detailed in Section 4). The engine maintains dimension scores, confidence levels, evidence counts, and context-variant weights.

Memory Store. Processed conversations are stored as derived summaries, topic tags, and extracted observations rather than raw transcripts. A three-tier lifecycle (active, archive, consolidated) manages memory over time.

Retrieval System. When context is requested, the system combines semantic vector search with cognitive science-informed re-ranking (detailed in Section 5) to surface the most relevant prior context.

Taste Subsystem. A distinct preference modeling layer tracks accumulated taste signals across domains, with confidence scoring and contradiction awareness (detailed in Section 6).

3.2 Privacy-Preserving Design

A foundational architectural decision is that the system does not store raw conversation transcripts. Each conversation is processed into structured outputs (cognitive scores, preference observations, topic summaries, entity tags) and the source messages are discarded. This design has significant implications for both privacy and evaluation, discussed in Sections 7 and 8.

4. Cognitive Fingerprinting Framework

4.1 Dimensions

The cognitive fingerprint models user behavior across ten dimensions, each representing a behavioral spectrum observable from natural conversation:

These dimensions were selected through iterative testing against production conversations, with three criteria: (1) each dimension must be reliably observable from conversational behavior without requiring self-report instruments, (2) each dimension must be orthogonal enough to carry independent information, and (3) each dimension must meaningfully influence how an AI system should interact with the user.

We deliberately avoid psychometric dimensions (e.g., Big Five personality traits, MBTI categories) that require validated self-report instruments. Our dimensions are behavioral, measuring what the user demonstrably does rather than what a personality model predicts they would do.

4.2 Bayesian Score Updating

Each dimension d is modeled as a continuous score θ_d on [-1, 1], where -1 represents the low end of the spectrum, +1 represents the high end, and 0 represents no observed tendency. The score is updated via a weighted Bayesian update rule:

This update rule has desirable properties: early observations with low n shift the score significantly (rapid initial learning), while later observations with high n produce smaller updates (stable profile), consistent with Bayesian posterior convergence.

This produces a confidence curve that rises quickly with initial observations and asymptotically approaches certainty, reflecting the diminishing marginal information value of additional consistent observations.

This ensures that the cognitive profile reflects current behavior rather than historical patterns that may no longer apply, while still retaining long-established tendencies at reduced weight.

4.3 Context-Variant Scoring

A critical insight from deployment is that users exhibit different cognitive patterns across contexts. A person may reason top-down in business settings but bottom-up when working on creative projects. They may prefer dense information in technical discussions but brief summaries for personal planning.

We maintain context-variant scores for each dimension across defined contexts (e.g., business, personal, creative, technical). When context is ambiguous, the system falls back to the general score. When context is clear, the context-specific variant takes precedence, with the general score serving as a Bayesian prior for contexts where insufficient evidence has accumulated.

5. Retrieval Augmentation via Cognitive Science Priors

5.1 Motivation

Standard retrieval in memory-augmented LLM systems relies on cosine similarity between the query embedding and stored memory embeddings. While effective for topically relevant retrieval, cosine similarity is agnostic to several factors that cognitive science identifies as central to human memory:

• Unresolved tasks occupy disproportionate mental space (Zeigarnik, 1927)
• Emotionally significant events are remembered more readily (McGaugh, 2004)
• Frequently accessed, recently active information surfaces more easily (Anderson & Schooler, 1991)
• Current mood influences which memories are accessible (Bower, 1981)

We apply these findings as retrieval augmentation, re-ranking results from the base semantic search.

5.2 Augmentation Framework

Given a base similarity score sim(q, m) for query q and memory m, the augmented retrieval score is:

Zeigarnik Boost (β_z). Memories tagged as unresolved (containing open questions, pending decisions, or incomplete tasks) receive elevated retrieval priority. This models the Zeigarnik effect, the well-documented phenomenon that incomplete tasks are recalled more readily than completed ones (Zeigarnik, 1927; Baumeister & Bushman, 2008).

Emotional Salience Boost (β_e). Memories associated with heightened emotional states (frustration, excitement, satisfaction, urgency) receive elevated priority. The emotional state is assessed at storage time and preserved as metadata. This models the established finding that emotional arousal enhances memory consolidation and retrieval (McGaugh, 2004; Kensinger, 2009).

Frequency-Recency Boost (β_f). Memories whose topic tags appear frequently across the user's conversation history, and especially those with recent activity, receive elevated priority. This models the frequency-recency interaction in human memory, where items that are both frequently encountered and recently accessed are most available (Anderson & Schooler, 1991).

Mood-Congruent Boost (β_m). When the current user mood can be inferred from the conversation, memories associated with congruent emotional states receive elevated priority. This models mood-congruent memory retrieval, where current emotional state facilitates access to memories encoded in similar states (Bower, 1981).

The multiplicative composition ensures that boosts compound (a memory that is both emotionally salient and recently relevant benefits from both factors) without any single factor dominating the ranking.

5.3 Results

We evaluated the augmented retrieval system against the base cosine-similarity retrieval on a held-out set of queries from production usage:

The improvements are consistent across query types, with the largest gains on queries involving ongoing projects (where Zeigarnik boosting is most relevant) and emotionally significant topics (where salience boosting contributes most).

6. Taste Modeling as Distinct Subsystem

6.1 Motivation

We identify taste as a category of user knowledge that is distinct from both memory (factual recall of past events) and cognitive profile (stable reasoning patterns). Taste encompasses accumulated preferences, aesthetic judgments, and domain-specific opinions that evolve over time, sometimes contradicting earlier positions.

A user's preference for minimalist design is not a fact to be retrieved (memory) nor a reasoning pattern to be modeled (cognitive profile). It is a taste signal that should inform recommendations, evaluations, and creative work. Treating taste as a separate subsystem allows for specialized handling of preference evolution, contradiction, and domain specificity.

6.2 Framework

The taste subsystem maintains preference records across domains, each with: the preference itself (expressed as a structured observation), a confidence score (reflecting evidence strength and recency), source type (distinguishing stated from revealed preferences), domain classification (enabling domain-specific retrieval), and temporal metadata (supporting staleness detection).

Preferences supported by multiple independent observations have higher confidence. Preferences not reinforced over extended periods decay in confidence. Preferences that have been contradicted by later behavior are marked and their confidence is reduced, but they are not deleted, as preference contradictions themselves carry information.

6.3 Stated vs. Revealed Preferences

A critical distinction in the taste system is between what users say they prefer and what their behavior reveals. When a user states "I prefer concise responses" but consistently engages more deeply with detailed, thorough explanations, the system maintains both signals and weights revealed preferences more heavily over time. This models the well-established gap between stated and revealed preferences in behavioral economics (Samuelson & Zeckhauser, 1988).

7. Evaluation

7.1 Critique of Existing Benchmarks

The primary benchmarks in use for AI memory systems, MemoryBench (Shah, 2025) and LoCoMo (Maharana et al., 2024), evaluate transcript retrieval: given a question about what was said in a conversation, can the system retrieve the relevant passage? This measures a genuine capability, but it is the wrong capability for evaluating cognitive memory systems.

A system designed to store derived cognitive signals rather than raw transcripts will necessarily score poorly on transcript retrieval benchmarks. Our system scores 20% on MemoryBench (2 of 10 queries correct, both adversarial). This is correct behavior. The system cannot retrieve transcripts it deliberately does not store.

The analogy is evaluating a summarization system on verbatim recall. A good summarizer discards surface text in favor of meaning. Penalizing it for not reproducing the original text measures the wrong objective.

7.2 PaoloBench

PaoloBench evaluates cognitive memory accuracy across seven categories: preference recall, decision pattern recognition, cross-session context linking, temporal reasoning (understanding when things happened relative to each other), cognitive profile accuracy, entity knowledge (remembering facts about people and projects in the user's life), and contradiction handling.

7.3 MemoryDelta

MemoryDelta measures whether cognitive memory actually improves AI interactions. The methodology is a blind A/B comparison: an LLM generates a response without cognitive context (baseline), then with the full cognitive profile injected (treatment). A judge model evaluates which response better serves the specific user.

We publish our MemoryBench score (20%) alongside our PaoloBench score (95.7%) because intellectual honesty about what a system does and does not do matters more than cherry-picking favorable metrics.

8. Privacy-Preserving Architecture

8.1 Design Philosophy

The system implements privacy preservation as an architectural constraint rather than a policy overlay. Raw conversation messages are processed into structured outputs and then discarded. The system literally cannot retrieve what it does not store.

8.2 Tradeoffs

Privacy guarantee. No raw user messages exist in the system after processing. A data breach would expose cognitive scores and topic summaries, not conversation transcripts.

Cognitive compression. Storing derived signals rather than raw text produces a more compact, more queryable representation. A hundred conversations about kitchen renovation compress to a small set of preference observations and cognitive signals, rather than megabytes of raw text.

Benchmark misalignment. As discussed in Section 7.1, the system scores poorly on retrieval-oriented benchmarks by design. This is an acceptable tradeoff given that the system optimizes for a different objective.

9. Extension to Organizational Agent Memory

9.1 From Individual to Organizational Cognition

As AI agents proliferate within organizations, each performing specialized tasks (inventory management, customer support, advertising, daily reporting), a new memory challenge emerges: how do agents serving the same organization share context without losing role-specific focus?

We extend the cognitive fingerprinting framework to multi-agent organizational memory. In this architecture, each agent maintains its own scoped memory (what it has observed and learned in its specific role) while also contributing to and querying a shared organizational knowledge layer.

9.2 Architecture

Agent-scoped memory (write). Each agent stores its findings under its own identity. An inventory agent's observations about stock patterns are attributed to that agent and primarily retrievable by that agent.

Agent-scoped recall (read). Each agent can search its own history, providing continuity across runs. An advertising agent can recall what campaigns it analyzed yesterday.

Organization-scoped context (read). Any agent can search across the entire organizational knowledge base. A daily digest agent can surface relevant findings from the inventory agent, customer support agent, and advertising agent to produce a comprehensive briefing.

This architecture solves the coordination problem that emerges when multiple agents operate independently: without shared memory, agents duplicate work, miss cross-functional insights, and fail to build on each other's findings. With role-scoped writes and organization-scoped reads, each agent contributes its specialized knowledge while any agent can access the collective intelligence.

9.3 Implications

The organizational memory layer suggests that cognitive infrastructure extends beyond individual cognitive modeling. Teams develop collective cognitive patterns, institutional knowledge that transcends any individual member, and shared context that enables coordinated action. Modeling this organizational cognition, where the human users and their AI agents share a common memory substrate, represents a significant direction for future work.

10. Discussion

10.1 Cognitive Infrastructure as a New Layer

We argue that cognitive infrastructure, the layer that models how users think and makes that understanding available to any AI system, represents a distinct and necessary addition to the AI stack. It is not a feature of the model layer (models are stateless), the orchestration layer (orchestration manages workflow, not understanding), or the data layer (databases store information, not cognition).

Cognitive infrastructure sits between the user and every AI system they interact with, providing continuity and personalization that no single platform can offer. As users interact with multiple AI assistants, coding tools, search engines, and autonomous agents, the need for a portable cognitive layer that travels with the user becomes acute.

10.2 Compounding Value

Unlike retrieval-based memory systems, where value is approximately linear with data volume, cognitive profiles exhibit compounding value. Each observation refines the profile, making it more accurate, which makes interactions more effective, which produces richer conversations, which generate better observations. This feedback loop creates increasing returns over time, a property that flat retrieval systems do not possess.

10.3 Future Directions

Cross-platform portability. Cognitive profiles could serve as a portable identity layer, allowing users to carry their cognitive context across AI platforms without rebuilding from scratch on each new system.

Agent-to-human cognitive alignment. Agents that maintain cognitive models of the humans they serve can proactively adapt their behavior, communication style, and decision-making to align with user expectations.

Organizational cognitive modeling. Extending beyond individual profiles to model team-level cognitive patterns, institutional decision-making styles, and collective knowledge structures.

10.4 Limitations

Several limitations should be noted. The cognitive dimensions were developed iteratively in a production system rather than derived from a comprehensive cognitive science taxonomy, and their completeness and orthogonality have not been formally validated. The evaluation frameworks (PaoloBench and MemoryDelta) are novel and have not been externally validated against established psychometric instruments. The system has been deployed with a limited user base and its performance characteristics at larger scale are not yet established. The Bayesian update rule, while effective in practice, makes simplifying assumptions about observation independence that may not hold across all dimension-context pairs.

11. Conclusion

We have presented a cognitive fingerprinting framework that models how users think rather than storing what they said. The framework operates across ten behavioral dimensions using Bayesian updating with temporal decay, augments retrieval with cognitive science priors, maintains a distinct taste modeling subsystem, and extends to multi-agent organizational memory.

Our evaluation demonstrates that cognitive modeling produces measurable personalization improvements (70% win rate, +3.65 lift) while our critique of existing benchmarks reveals that the dominant evaluation paradigm measures the wrong capability for this class of system.

The AI memory field is moving rapidly, with over $40 million in venture funding deployed to memory-as-retrieval systems in the past six months alone. We believe the next significant advance lies not in better retrieval but in genuine cognitive understanding, and we have presented both the theoretical framework and production evidence for this claim.