---
mw_bundle: 1
id: wa-K_2rF
title: "AI Memory Research"
url: https://memory.wiki/b/wa-K_2rF
document_count: 3
updated: 2026-05-24T17:26:33.037Z
analysis_generated_at: 2026-05-14T19:20:02.680Z
source: "memory.wiki"
---
# AI Memory Research

> Captured conversations + external reading on how AI memory architectures actually work. Reading order: the Claude conversation lays out the three patterns, the Karpathy summary names the philosophical frame, the formatting tour is a side-quest reference.

**Intent:** Decide how mdfy's personal memory layer should compose with other AI tools.

## Summary

These documents establish the philosophical and technical foundation for mdfy's personal memory layer, arguing that human-authored memory beats AI-inferred memory across three architecture patterns. They validate mdfy's hub-shaped approach through comparison with vector recall and episodic snapshot alternatives, while demonstrating the content formatting capabilities that make URL-addressable knowledge viable.

## Themes

- human-authored memory superiority
- composable architecture design
- memory as curation problem
- URL-addressable knowledge systems

## Cross-document insights

- The battle between memory architectures isn't about retrieval efficiency, it's about who controls the narrative: humans who curate deliberately vs. AIs that infer from conversation fragments.
- mdfy's three-scope architecture (doc/bundle/hub) solves a problem Karpathy's single-wiki concept couldn't: providing project-specific context without requiring folder discipline from users.
- The formatting tour document reveals mdfy's strategic bet: by handling rich content formats natively, it removes friction from the human authoring process that competing memory systems create.
- All three documents converge on the same actionable principle: memory systems should optimize for human editing and curation, not just AI retrieval convenience.

## Key takeaways

- Human-authored memory architectures outperform AI-inferred memory because humans control the narrative and can edit/curate their knowledge over time.
- mdfy's composable three-scope design (doc/bundle/hub) provides better project-specific context than single-wiki approaches without requiring user discipline.
- The technical capability to handle rich content formats is crucial for reducing friction in the human authoring process that makes memory systems successful.

## Open questions / gaps

- No discussion of how mdfy should handle real-time collaboration or multi-user scenarios when composing with AI tools.
- Missing analysis of how mdfy's memory layer should integrate with AI tools that have their own memory systems (conflict resolution, synchronization).
- No exploration of privacy and access control when AI tools read from mdfy's URL-addressable knowledge.
- Lack of concrete examples showing mdfy composing with specific AI tools in real workflows.

## Notable connections

- **doc:T7ZGdpOm** ↔ **doc:yt2ZZqyI** — Karpathy's conceptual framework gets validated through Claude's comparative analysis of memory architectures.
- **doc:yt2ZZqyI** ↔ **doc:F1q9U1YP** — Claude's theoretical validation of hub-shaped memory is made practical through mdfy's rich content formatting capabilities.
- **doc:T7ZGdpOm** ↔ **doc:F1q9U1YP** — Karpathy's human-authored memory principle requires the technical foundation that mdfy's formatting capabilities provide.

## Concepts (this bundle)

- **Human-authored memory**
- **Composable scopes**
- **Vector recall**
- **Episodic snapshots**
- **Hub-shaped memory**
- **URL-addressable knowledge**
- **Curation vs retrieval**

## Concept relations

- **Human-authored memory** ↔ **Curation vs retrieval** — supports framework
- **Hub-shaped memory** ↔ **Human-authored memory** — implements principle
- **Composable scopes** ↔ **Hub-shaped memory** — extends architecture
- **URL-addressable knowledge** ↔ **Hub-shaped memory** — enables approach

## Documents

### 1. [AI memory architectures: a Claude conversation](https://memory.wiki/yt2ZZqyI)
> The captured Claude conversation. Read first — it names the three architectures.
Captured from a working session with Claude Opus, 2026-03-12. Cleaned, structured, and saved as a permanent URL so the next AI session can pick up where we left off.
*sections:* The question: What architecture should a personal memory layer use? Three patterns are in production today: | What Claude argued: > Vector recall trades precision for breadth. Episodic snapshots trade verbosity for fidelity. Hub-shaped memory trades automation for author-control. | Takeaway for mdfy: This is the existing direction. Worth checking against the spec page — /spec already documents the URL contract. No code change needed; this conversation just v | Related concepts: Vector recall, episodic snapshot, hub-shaped memory; Forgetting as a feature; URL-addressable knowledge

### 2. [Karpathy's LLM Wiki concept, in one read](https://memory.wiki/T7ZGdpOm)
> Karpathy's philosophical frame. Same conclusion, opposite direction.
Andrej Karpathy, Twitter thread (2024). Topic: a personal LLM-readable wiki — one place a person writes their knowledge, the LLM reads it instead of building memory by inference.
*sections:* Source: Andrej Karpathy, Twitter thread (2024). Topic: a personal LLM-readable wiki — one place a person writes their knowledge, the LLM reads it instead of building me | Core argument: > The most reliable AI memory is the one the human authored. Anything inferred from a chat transcript is lossy; anything written deliberately is durable. | Where Karpathy's vision stops: Single unified wiki. One person, one wiki. The structure is whatever the user makes inside that wiki. | Where mdfy goes further: Three composable scopes instead of one: doc, bundle, hub. The same URL primitive scales from a one-note share to a project context to a full knowledge graph. Ka | …

### 3. [Formatting tour: math, diagrams, code, tables](https://memory.wiki/F1q9U1YP)
> Reference for the rendering vocabulary used across the hub.
A reference for what renders in mdfy. Every block below appears in real docs across this hub.
*sections:* KaTeX math: Inline: $E = mc^2$. Display: | Mermaid diagrams: mermaid | Code with highlighting: typescript | Tables: | Scope | URL | Cost |


_Digest view — follow any link above to fetch that doc's full markdown. Add `?full=1` to this URL for the concatenated payload._