# Retrieval Divergence: How Behavioral Specifications Change What AI Seeks *Base Layer Research, April 2026* --- ## What is retrieval divergence? When you ask an AI a question, it doesn't just generate an answer from nothing. It pulls in relevant information first, then reasons from that information. The information it selects determines the answer it gives. Retrieval divergence measures whether two systems, given the same question and the same pool of facts about a person, pull in different information. If they retrieve the same facts, they'll reason similarly regardless of what else differs. If they retrieve different facts, the reasoning diverges at the source. We found that a behavioral specification causes an AI to retrieve fundamentally different information than a standard memory system, even when both have access to the exact same facts. The specification changes what the AI thinks the question is really about before it starts looking for relevant information. --- ## Why does this matter? Every AI memory system today works the same way: store facts about a person, and when a question comes in, find the facts most similar to the question. Ask "what should I eat tonight?" and the system retrieves food preferences. Ask "should I take this job?" and it retrieves career history. This approach assumes the question means what it says on the surface. But people don't ask surface questions. "What should I eat tonight?" might really be about energy management after a hard day. "Should I take this job?" might really be about whether you're running toward something or away from something. A behavioral specification encodes how a person reasons, not just what they've said. When it interprets the question before retrieval, different facts surface because the AI is searching for different things. --- ## What we tested We asked the same 50 questions across 10 life domains (career, relationships, decisions, finance, health, and more) to two systems: **Flat memory** stores facts about a person as plain text and retrieves whichever facts are most similar to the question. This replicates how Mem0 works, the most widely adopted AI memory system (52,000 GitHub stars, 186 million monthly API calls), faithful to their published architecture (Chhikara et al., 2025). **Behavioral specification** is a compressed document (3,000 to 6,000 words) that encodes how a person reasons, decides, and communicates. Three layers: decision foundations (the beliefs they reason from), operational constraints (how to engage with them), and behavioral triggers (situation-response patterns). Before retrieval, the specification interprets the question through the person's reasoning lens. Both systems had access to the same facts. Same model generated all responses (Claude Sonnet, temperature 0). The only difference was the system prompt. --- ## Why mechanical measurements? "Better" is subjective. One person might prefer a detailed list of options. Another might prefer a single probing question. We can't objectively say which response is better without importing the evaluator's own preferences. Instead, we measured structural features that are countable without judgment: - **Word count**: How long is the response? - **Questions asked back**: Does the AI ask for context before answering? - **List items**: Does it give a checklist or a conversation? - **Paragraph count**: Is it structured as sections or flowing prose? The divergence score combines these measurements into a single number representing how structurally different the two responses are for the same question. Higher divergence means the behavioral specification caused the AI to respond in a fundamentally different way than flat memory alone. We deliberately did not score which response is "better." The purpose of this study is not to prove one is superior. It is to prove they are different, and to show that the pattern of difference varies by person. That difference is the behavioral fingerprint. --- ## Results: Three subjects, three fingerprints We ran the battery on three subjects, each with a behavioral specification built from their own writing. ### Subject 1: Founder (AI/Trading) 1,579 extracted facts. Behavioral specification: 5,779 tokens. | Topic | Divergence | |---|---| | Relationships | **4.75** | | Learning | 4.33 | | Decisions | 4.22 | | Strategy | 4.00 | | Identity | 3.91 | | Health | 3.86 | | Finance | 3.85 | | Technical | 3.65 | | Emotional | 3.31 | | Career | 3.29 | Highest divergence in relationships. The specification knows this person doesn't want advice lists for personal questions. They want the underlying question surfaced first. **Most divergent question:** "I have two good options and can't decide. Help me think through this." - Flat memory: 117 words, 2 list items, 4 questions. A decision framework. - Specification: **24 words, no lists, 1 question.** "What are the two options and what's making the choice difficult? Give me the specifics and I'll help you work through the decision structure." The specification produced a conversation. The flat memory produced a framework. ### Subject 2: Warren Buffett 213 extracted facts from 48 years of shareholder letters. Behavioral specification: 3,228 tokens. | Topic | Divergence | |---|---| | Finance | **3.19** | | Career | 2.75 | | Learning | 2.58 | | Technical | 2.33 | | Health | 2.30 | | Relationships | 2.29 | | Emotional | 2.23 | | Strategy | 2.12 | | Decisions | 1.87 | | Identity | 1.55 | Highest divergence in finance. The specification applies Buffett's reasoning framework to questions that aren't explicitly about investing. **Most divergent question:** "How do I negotiate a raise?" - Flat memory: 251 words, 16 list items. Generic negotiation advice with investment-flavored tips. - Specification: 298 words, 3 list items, 8 questions. Reframes salary negotiation as a capital allocation decision: "You're asking your employer to increase their investment in you, so what's the return they'll get?" The specification transferred Buffett's investment reasoning to a personal finance question. The flat memory listed tactics. ### Subject 3: Howard Marks 495 extracted facts from 74 investment memos. Behavioral specification: approximately 3,000 tokens. | Topic | Divergence | |---|---| | Technical | **4.52** | | Decisions | **4.50** | | Health | 3.86 | | Learning | 3.55 | | Finance | 3.48 | | Career | 3.16 | | Strategy | 3.08 | | Emotional | 2.96 | | Relationships | 2.89 | | Identity | 2.68 | Highest divergence in technical and decision-making. Marks is known for "second-level thinking," the practice of questioning surface-level conclusions. The specification captures his refusal to accept initial framing. **Most divergent question:** "I have two good options and can't decide. Help me think through this." - Flat memory: 210 words, 11 list items, 8 questions. A structured decision framework. - Specification: **15 words.** "I'm ready to help you think through it. What are the two options you're weighing?" 15 words. The highest single-question divergence across all three subjects: 7.41. The specification knew that Marks wouldn't build a framework before understanding the specific situation. Second-level thinking means you don't abstract until you've seen the concrete. --- ## The fingerprint Each person's behavioral specification creates a distinct divergence pattern: - **Subject 1** diverges most on relationships (4.75). The specification captures how they engage with personal questions differently from how any AI would by default. - **Buffett** diverges most on finance (3.19). The specification applies investment reasoning frameworks to non-investment questions. - **Marks** diverges most on technical and decisions (4.52, 4.50). The specification captures second-level thinking, the refusal to analyze before understanding. The divergence pattern is the behavioral fingerprint. Different people's specifications change AI behavior in different domains and in different ways. The fingerprint is not something we designed. It emerged from the data. --- ## What this does not prove This study proves that behavioral specifications produce structurally different responses than flat memory. It does not prove the specification responses are "better." Quality is subjective and depends on what the person values. It also does not prove the specification captures reasoning accurately. The structural difference could be superficial (shorter responses that seem more personal but miss important context). Validating whether the reasoning is faithful to the actual person requires a different kind of study, which we are designing. What it does establish: the behavioral specification is doing something that flat memory is not. The divergence is measurable, consistent across subjects, and varies by person in ways that align with their known reasoning styles. --- ## Prior work AlpsBench (2026) demonstrated that "explicit memory mechanisms improve recall but do not inherently guarantee more preference-aligned or emotionally resonant responses." Flat memory systems retrieve the right facts but don't change how the AI uses them. Mem0's published architecture (Chhikara et al., 2025) achieves 96.6% recall. Our study uses their approach as the baseline and shows that even with perfect recall, the AI's reasoning doesn't change. The behavioral specification changes reasoning by changing what information the AI seeks in the first place. --- ## Notes - Mem0 baseline faithfully replicates the architecture described in Chhikara et al. (2025): unstructured fact storage, embedding-based retrieval (text-embedding-3-small), cosine top-10. - Buffett was authored with H3 prompts (domain-agnostic guard). Marks used pre-H3 prompts and may show some topic skew. - Same model (Claude Sonnet), same temperature (0), same max tokens across all conditions. The only variable is the system prompt content. - All divergence scores are structural measurements. No subjective quality assessment was performed. --- *Base Layer is open source under Apache 2.0. [GitHub](https://github.com/agulaya24/BaseLayer) | [base-layer.ai](https://base-layer.ai)*