The Meta-Narrative Problem
What Search Gets Wrong
When you search your old emails, journal entries, or social media posts, you're asking: "What did I write about X?"
But the more interesting question is: "How have my views on X changed over time?"
No existing tool answers this. Not Google, not Notion, not Obsidian, not any digital memory system. They're all built around the same paradigm: documents as independent retrieval targets.
I call this gap the meta-narrative problem: the inability to understand the story your data tells about your personal evolution.
Why It Matters
Consider these scenarios:
- You journaled about "valuing independence" in 2020. By 2024, you're writing about "the importance of community." Neither entry references the other. No search query would surface this shift.
- Your Reddit comments in 2021 were skeptical of remote work. Your LinkedIn posts in 2023 celebrate it. These live on different platforms with different exports. No tool connects them.
- Your Obsidian notes from 2022 express confidence in a career direction. Your WhatsApp messages from 2023 reveal uncertainty about the same topic. No system correlates beliefs across platforms.
The Four Requirements
Solving the meta-narrative problem requires four capabilities that no existing tool combines:
- Multi-platform data fusion — Ingest heterogeneous exports from dozens of platforms into a single temporal knowledge base
- Belief extraction — Move beyond keyword matching to identify structured beliefs, values, and self-descriptions from unstructured text
- Temporal contradiction detection — Compare beliefs across time to find genuine shifts, not just topic overlap
- Narrative synthesis — Generate human-readable stories about how thinking evolved, grounded in original writing
How LLMs Change the Game
Before LLMs, each of these steps required task-specific training. Extracting beliefs from text needed a custom NLP model. Generating coherent narratives needed another. The engineering cost was prohibitive for a personal tool.
Modern LLMs can do all four zero-shot. The challenge shifts from "can we do this?" to "can we do this efficiently and privately?"
MemryLab's answer: embedding-based pre-filtering reduces LLM calls from O(n^2) to O(k), where k <= 20. The entire pipeline runs in ~25 minutes on a consumer laptop with a local 8B model.
The Bigger Picture
The meta-narrative problem isn't just a technical curiosity. It's about self-knowledge — the kind that only comes from seeing patterns you couldn't see in the moment.
When MemryLab surfaces a contradiction between your 2021 self and your 2024 self, it's not pointing out a flaw. It's showing you growth. And that's worth building for.