Current personalization and memory architectures in AI applications are strictly additive: they profile what a user does or says, treating memory as a compression task.
However, they fail to capture the negative space -- the critical actions a user is not taking. To evolve from a reactive tool into a proactive coach, AI systems must recognize when the absence of an action implies a gap in the user's trajectory.
Standard memory pipelines are reactive by design. They wait for user input, summarize it, and store it. By only looking at the positive space, the model remains trapped in a feedback loop of what is already known.
If a user is learning Japanese but never asks about listening comprehension, a standard "additive" memory simply records that the user is good at Kanji. It doesn't realize that the user is becoming "deaf" to the spoken language. To fix this, we need a pipeline that performs Shadow Profiling: generating hypotheses about what the user should be doing but isn't.
While this thesis could potentially be applied to many use cases, for now let's focus on one concrete CUJ -- providing coaching to users going through Growth Periods: longitudinal, goal-oriented phases with established best practices.
Throughout this exploration, we'll follow three archetypical users:
We can implement this through a four-step offline pipeline that runs periodically.
The pipeline scans recent logs to detect if the user has entered a sustained project. It reconstructs a Latent Spec -- the inferred goal behind the queries.
| User | Signal | Latent Spec |
|---|---|---|
| Language Student | 4 weeks of queries on kanji and grammar | Intermediate Japanese self-study |
| Solo Developer | Queries about React, Tailwind, and Vercel | Solo dev building a web app |
| Wannabe Athlete | Logs of progressively longer runs | Novice runner training for a race |
The pipeline uses a frontier model to dynamically generate a "success checklist" or Competency Tree for that specific Latent Spec. This is the benchmark of what a "complete" version of this journey looks like.
| User | Competency Tree |
|---|---|
| Language Student | Reading, Writing, Listening, Speaking, Culture |
| Solo Developer | Frontend, Data Fetching, Database Schema, Auth, Deployment |
| Wannabe Athlete | Base Mileage, Speedwork, Recovery, Nutrition, Tapering |
The pipeline uses a frontier model to look for gaps between the user's actual activities and the Competency Tree.
| User | Observed Activity | The Gap |
|---|---|---|
| Language Student | High coverage on Reading/Writing | Zero activity in Listening/Speaking |
| Solo Developer | High coverage on Frontend/Deployment | Zero mentions of Route Protection or Auth |
| Wannabe Athlete | High coverage on Mileage/Gear | Zero mentions of Intra-run Nutrition (electrolytes/gels) despite hitting 15-mile runs |
The pipeline outputs these gaps as Shadow Memories. These are injected into the context for future sessions as "soft constraints".
| User | User Query | AI Intervention |
|---|---|---|
| Language Student | "Give me a list of N3 vocabulary." | "Here is the list. Since you've been doing great with reading lately, do you want to try an audio exercise with these words to practice listening comprehension?" |
| Solo Developer | "How do I trigger this production build?" | "Here is the build command. Before we push, I noticed we haven't discussed locking down the API routes—do you need a quick drop-in Auth solution to secure this first?" |
| Wannabe Athlete | "Logged 16 miles today. Legs are heavy." | "Great milestone! Since you are crossing that 15-mile threshold, you'll start depleting glycogen stores rapidly. Have you started testing a fueling strategy for these longer efforts?" |
By systematically profiling the negative space, we shift the AI's role from a passive responder to a high-context partner. It allows us to guide users around pitfalls they don't even know exist.
The future of AI memory isn't just about remembering what we said—it's about understanding what we've forgotten to do.
Generative AI Usage Disclosure: This post is first written by me after a long discussion with Gemini 3.0 Pro and then refined by it.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。