---
mw_bundle: 1
id: wpwVCSDF
title: "AI memory research: the public frontier"
url: https://memory.wiki/b/wpwVCSDF
document_count: 4
updated: 2026-05-14T19:20:10.811Z
analysis_generated_at: 2026-05-14T19:20:10.811Z
source: "memory.wiki"
---
# AI memory research: the public frontier

> Side-by-side notes on Mem0, Letta, Microsoft GraphRAG, Karpathy's LLM Wiki, llms.txt adoption.

**Intent:** Curated bundle: AI memory research: the public frontier

## Summary

This collection explores the current frontier of AI memory research from a practical implementation perspective, comparing automated extraction systems (Mem0/Letta) with user-authored approaches (mdfy/Obsidian wikis), examining Microsoft's GraphRAG breakthrough in knowledge graph reasoning, and tracking emerging standards like llms.txt for AI discoverability. The documents collectively map the tension between AI-driven vs human-curated knowledge management and the various technical approaches being pursued in the field.

## Themes

- extracted vs authored memory
- knowledge graph reasoning
- AI discoverability standards
- personal knowledge management

## Cross-document insights

- The fundamental architectural divide in AI memory is not about technology but about authorship: who creates the knowledge representation that the AI consumes - the AI itself through extraction, or the human through curation.
- GraphRAG's community detection approach suggests that graph topology matters more than embedding similarity for complex reasoning tasks, potentially making structured knowledge graphs superior to vector databases for multi-hop queries.
- The early adoption pattern of llms.txt reveals that AI memory standards are being driven by developer tool companies who have both the technical capability and direct user feedback loops, not by AI research labs.
- The convergence on wiki-shaped knowledge (interconnected markdown pages) across multiple systems suggests this may be the optimal format for human-AI knowledge collaboration, regardless of the underlying technical implementation.

## Key takeaways

- AI memory research is splitting into two distinct paradigms: automated extraction systems that infer user profiles from behavior, and user-authored systems that rely on human curation of knowledge.
- GraphRAG represents a significant advance in AI reasoning capabilities by using graph structure and community detection rather than just semantic similarity for information retrieval.
- The practical adoption of AI memory standards is being led by developer tool companies who have direct incentives and technical capability, suggesting a bottom-up rather than top-down evolution of the field.

## Open questions / gaps

- Quantitative benchmarks comparing extracted vs authored memory systems on standardized tasks - all comparisons are currently qualitative and anecdotal.
- Analysis of hybrid approaches that combine automated extraction with human curation, which may represent the optimal path forward.
- Privacy and security considerations for AI memory systems, especially regarding data persistence and cross-system sharing.
- Cost-benefit analysis of different memory approaches at scale, including infrastructure and maintenance costs over time.

## Notable connections

- **doc:_ybJOqIB** ↔ **doc:glqi_Xjw** — Both contrast automated AI extraction with user-authored approaches, with mdfy appearing as a user-authored alternative in both documents.
- **doc:6WkjlKgA** ↔ **doc:_ybJOqIB** — GraphRAG's knowledge graph approach represents a more sophisticated backend for memory systems compared to the simpler extraction methods used by Mem0 and Letta.
- **doc:glqi_Xjw** ↔ **doc:qKRfGtCa** — Karpathy's wiki vision and llms.txt both address the challenge of making personal knowledge accessible to AI systems, but through different technical mechanisms.
- **doc:6WkjlKgA** ↔ **doc:qKRfGtCa** — GraphRAG's service-based approach contrasts with llms.txt's URL-based discoverability model, representing two different paradigms for AI knowledge access.

## Concepts (this bundle)

- **Extracted Memory**
- **User-Authored Memory**
- **Knowledge Graphs**
- **Community Detection**
- **Multi-hop Reasoning**
- **AI Discoverability**

## Concept relations

- **Extracted Memory** ↔ **User-Authored Memory** — contrasts with
- **Knowledge Graphs** ↔ **Multi-hop Reasoning** — enables
- **Community Detection** ↔ **Knowledge Graphs** — analyzes structure

## Documents

### 1. [Microsoft GraphRAG: what we learned](https://memory.wiki/6WkjlKgA)
Read the 2024 paper and the follow-ups (Project NotebookLM, the v1.1 release, the open-source community fork). Notes for the team to reference when "GraphRAG" comes up.
*sections:* The thesis, in one sentence: > Build a knowledge graph from a document corpus, run community detection on it, and answer queries by traversing the graph instead of just embedding-matching. | What's good: Multi-hop reasoning. GraphRAG beats naive RAG on the "compare X and Y across documents" class of questions. The community-detection step gives it a structural p; Honest about its costs. The paper is explicit that GraphRAG is 10-50x more expensive than naive RAG to build, because the graph extraction is an LLM-per-chunk o; The open-source release is real. The Python package works. The community fork (rust-graphrag) shaves indexing time by ~3x. | …

### 2. [Mem0 vs Letta: extracted memory comparison](https://memory.wiki/_ybJOqIB)
Side-by-side after 3 weeks of using both in parallel on the same chat corpus.
*sections:* Setup: I forked my Claude Code session log (about 280 conversations, 6 months) and fed it through both Mem0 and Letta. Both got the same input, same time window. I ask | What Mem0 produced: Facts. Direct, dry, accurate. "User is building mdfy.app. User prefers Rust + TypeScript. User has shipped 3 Chrome extensions." Mostly nouns and verbs.; Extraction quality. Strong. It picked up the cross-AI thesis from the corpus correctly, including the specific phrase "structural moat."; Misses. Anything stylistic. It noted "user works in Korean and English" but didn't capture how I work in each. The where-and-when-I-write-Korean nuance was invi | …

### 3. [llms.txt adoption: who's actually shipping it](https://memory.wiki/qKRfGtCa)
Snapshot 2026-05-01. I'll re-run this every quarter.
*sections:* What llms.txt is, briefly: A plain-text discoverability file at the root of a site (/llms.txt) that tells AI agents what's available, in what shape, and where to fetch it. Inspired by rob | …

### 4. [Karpathy wiki: the parts that map](https://memory.wiki/glqi_Xjw)
Expansion of demo-karpathy-llm-wiki. That doc is the public summary; this is the internal "where it lines up with us and where it doesn't" working note.
*sections:* The Karpathy quote that started this: > "Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase." | What maps directly: The user is the author. Karpathy explicitly says the wiki is hand-curated. The AI is the programmer reading it, not the writer.; Markdown is the substrate. Every example he gave is markdown.; The "wiki" word. Both his framing and ours land on it. | What doesn't map: Three structural differences: | Where I think Karpathy is right: The deepest claim — that a personal knowledge layer is the missing piece between you and the AI tools you use every day — is exactly right. The disagreement is  | …


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