Shuichiro Ogawa
日本語

Notes · updated 2026-07-06

Knowledge Management Methods for the LLM Era (Mid-2026 Status Report)

How should knowledge be delivered to LLMs? The answer to this question has been rapidly systematized between late 2024 and mid-2026. What began as prompt engineering practice has progressed through context design (context engineering) to arrive at pre-compiled, persistently managed knowledge. This note surveys the methods that are converging as of the time of investigation, examining both their lineage and their structure.

Lineage

2024-11  MCP announced (Anthropic)
2025-04  Karpathy LLM Wiki gist published
2025-06  Tobi Lutke coins "context engineering"
2025-07  LangChain 4-operation taxonomy (Write/Select/Compress/Isolate)
2025-08  Addy Osmani / O'Reilly Radar systematization
2025-08  OpenAI publishes AGENTS.md
2025-09  Anthropic context engineering official blog post
2025-09  Claude Memory introduced (Pro)
2025-12  MCP + AGENTS.md donated to Linux Foundation AAIF
2026-01  MCP Apps released
2026-02  ETH Zurich AGENTbench (human curation > LLM auto-generation)
2026-03  Claude Memory rolled out to all accounts
2026-06  Google OKF v0.1 announced (formalization of LLM Wiki)
2026-06  NotebookLM agentic research update

1. Pre-Compiling Knowledge: Karpathy LLM Wiki to Google OKF

The Karpathy LLM Wiki Pattern

In April 2025, Andrej Karpathy published a design on GitHub Gist (5,000+ stars) built around a three-layer structure.

  • Raw layer (raw/): Immutable primary sources (papers, articles, images). The LLM reads but never writes.
  • Wiki layer (wiki/): A corpus of Markdown pages owned by the LLM. The LLM performs summarization, cross-linking, and consistency maintenance. index.md (category-based catalog) and log.md (append-only operation log) are reserved filenames.
  • Schema layer (CLAUDE.md, etc.): Configuration that encodes the wiki’s structure, naming conventions, and workflows. This transforms the LLM from a “general-purpose chatbot” into a “disciplined wiki maintainer.”

There are three operations. Ingest (read new raw material and update wiki pages), Query (search the wiki to answer questions, writing valuable answers back to the wiki), and Lint (detect contradictions, orphan pages, and broken links).

The core design insight is that the tedious part of maintaining a knowledge base is not reading or thinking but bookkeeping, and LLMs reduce that bookkeeping cost to near zero. Whereas RAG re-derives knowledge from raw chunks on every query, the LLM Wiki pre-compiles and accumulates knowledge.

Practitioner Application: Obsidian + Claude Code

In July 2026, Machina, an AI agency operator, published a long-form article on X (1.55 million views) demonstrating an implementation of the Karpathy pattern. The folder structure comprises four layers: raw/ (immutable primary sources), entities/ (one page per entity), concepts/ (one page per concept), and INDEX.md (a listing of all pages with one-line descriptions). The system uses an Obsidian vault as storage and Claude Code as the operating agent.

The operational rules fit in four lines. Each file holds a single lesson with a one-line summary at the top. Duplicates are avoided by updating existing pages. Notes found to be incorrect are deleted. Raw material and compiled pages are always kept separate. The underlying philosophy is identical to Karpathy’s original design, though the separation of entities/ and concepts/ represents a variant that subdivides Karpathy’s wiki/ by purpose.

Maintenance is automated through four loops. A session-end hook extracts insights and updates the wiki. A nightly compile pass uses a cheap model to process the day’s raw material. A weekly lint pass detects contradictions, duplicate pages, and dead links. Only the weekly synthesis pass uses a premium model, reading through the entire vault and documenting changes. The model-tiering principle is to delegate routine work to inexpensive models and limit the premium model to a single whole-vault integration pass (model-tiering patterns).

Research uses a fan-out pipeline. A single question is split into 3-5 sub-questions, and parallel agents each explore a different source surface (social media, official documentation, videos, etc.). Findings are recorded as receipts containing a claim, source link, and date; a skeptic agent then attacks each claim for verification. Claims supported by only a single source are rejected. Only findings that survive are committed to the vault with dates, links, and expiration dates.

The cost management design is based on a metaphor that treats the context window as “a room you pay admission to enter.” CLAUDE.md is automatically loaded at session start, incurring no additional cost. All other pages use a pay-per-read approach, navigated via links from INDEX.md or found via grep. When large-scale reading is required, a sub-agent processes it in a separate context and returns only a summary.

The tool stack mentioned includes ScrapeCreators (cross-platform retrieval of social media posts from the last 30 days), X MCP (live posts and bookmarks), yt-dlp (YouTube transcripts), Perplexity deep research (cited long-form research), and Firecrawl (full-page Markdown conversion). However, many of these are paid tools commercially endorsed by the author, and no methodological validation is provided.

Performance claims include an example where accuracy on an accounting task improved from 70% without client history to over 85-90% with history added, and an example where a mid-tier model with a voice profile outperformed Fable 5 without a profile. However, methodologies are not disclosed in either case [requires primary verification]. An Anthropic internal test is cited in which Fable improved round-over-round in a deck-building game with file-based memory, but this is a vendor test with no external replication.

Google OKF (Open Knowledge Format)

On June 13, 2026, Google Cloud released OKF v0.1, a vendor-neutral specification standardizing the Karpathy pattern.

  • bundle = directory tree (distributable as a git repository, tarball, or zip)
  • Each concept = a single .md file (file path serves as ID)
  • Reserved filenames: index.md (table of contents), log.md (change log)
  • Required frontmatter fields: type only. Recommended: title, description, resource (external URI), tags, timestamp
  • Extension fields are unrestricted; consumers must preserve unknown keys and must not reject them
  • Links: bundle-relative absolute paths such as /tables/orders.md are recommended. Links to nonexistent targets are valid

OKF separates producers (humans, pipelines, LLMs) from consumers (agents, viewers, other LLMs). Reference implementations ship with the spec, including an Enrichment Agent that walks BigQuery to auto-generate OKF, a Static HTML Visualizer for graph views, and sample bundles for GA4, StackOverflow, and Bitcoin.

2. Context Engineering: Designing the Context

Emergence of the Concept

In June 2025, Shopify CEO Tobi Lutke proposed context engineering on X (for the lineage of this vocabulary, see ai-engineering-taxonomy-design), defining it as “the art of providing all the context necessary for a task to be solvable by the LLM.” Karpathy endorsed the idea, describing it as “the delicate art and science of filling the context window with just the right information for the next step.”

In September 2025, Anthropic formalized the concept in the official blog post “Effective context engineering for AI agents,” defining it as “a set of strategies for curating and maintaining the optimal set of tokens at LLM inference time.” The post exceeded 500,000 page views.

Anthropic’s Four-Element Framework

  1. System Prompts: Write at the “right altitude.” Specific but flexible heuristics.
  2. Tools: Design to return token-efficient information. Avoid functional overlap.
  3. Examples (Few-shot): Diverse canonical examples rather than exhaustive edge cases.
  4. Message History & Runtime Retrieval: Just-in-time context. Dynamic loading via lightweight IDs.

Three techniques serve long-running tasks: Compaction (summary compression), Structured Note-taking (external memory such as NOTES.md), and Sub-agent Architecture (context isolation through sub-agents).

LangChain’s Four-Operation Taxonomy

LangChain classified context engineering into four operations by analogy with OS memory management.

OperationMeaningOS Analogy
WritePersist information outside the contextsave to disk
SelectLoad relevant information into the contextpage in
CompressReduce tokens (summarization, trimming)swap
IsolatePartition the context (multi-agent)process isolation

Human Curation vs. LLM Auto-Generation

The AGENTbench study from ETH Zurich (February 2026) reported that LLM-auto-generated context files reduced task success rates by approximately 3% while increasing costs by over 20%, whereas human-authored files improved success rates by approximately 4%. Manual curation by humans, combined with keeping files concise, is effective at least at the present stage.

3. Convergence of In-Repo Knowledge Files

As of mid-2026, five specifications have converged on nearly identical patterns (Markdown files at the repository root).

FileProviderScope ControlCharacteristics
AGENTS.mdOpenAI -> AAIFDirectory hierarchyOpen standard. 60,000+ repos adopted. Read by 6+ tools
CLAUDE.mdAnthropic4 layers (Managed/User/Project/Local)Auto Memory. @path import (4-hop recursion). /init auto-generation
.cursor/rules/*.mdcCursorYAML glob + 4 modesAlways/Intelligent/Specific Files/Manual application modes
.github/copilot-instructions.mdGitHubglob + directoryPersonal > Repository > Organization priority
.gemini/Google—For Gemini CLI

Cline adopts its own Memory Bank pattern (memory-bank/ containing 6 fixed Markdown files), with a convention of reading all files at the start of every task. Windsurf (now Devin Desktop) separates Rules (explicit instructions with 4 trigger modes) from Memories (auto-learned, locally stored).

Three design decisions are shared across these systems. (1) Markdown is the format. (2) File paths control scope. (3) Human-written instructions and LLM-written learning notes are kept separate.

4. Memory Systems

Claude Memory

Introduced for Pro users in September 2025 and rolled out to all accounts in March 2026. It takes a transparent, Markdown-file-based approach rather than using a vector database. A 39% lift in internal agentic search evaluation has been reported through the combination of memory and context editing [methodology verification needed].

Claude Code’s Memory comprises two channels: CLAUDE.md (human -> Claude) and Auto Memory (Claude -> Claude). Auto Memory is stored in ~/.claude/projects/<project>/memory/, with the first 200 lines of MEMORY.md loaded at session start.

ChatGPT Memory

OpenAI provides Projects (workspaces that group chats, files, and custom instructions by topic). In 2026, they began US rollout of Dreaming Memory (a new architecture that synthesizes useful context from past conversations) for Plus/Pro users.

Common Design Principles

Memory systems share three characteristics: (1) separation of explicit memory (human-written rules) from implicit memory (auto-written LLM learning outcomes), (2) file-based transparency (users can read and write, and content can be managed with git), and (3) a combination of automatic loading at session start with on-demand reference when needed.

5. MCP: The Protocol Layer for Tool Connectivity

Announced by Anthropic in November 2024 and donated to the Agentic AI Foundation (AAIF) under the Linux Foundation in December 2025. Adopted by OpenAI, Google DeepMind, and Microsoft, with Python + TypeScript SDKs reaching approximately 97 million downloads per month.

MCP Apps, released in January 2026, added the ability for tools to return HTML UIs rendered in sandboxed iframes. This works in Claude, ChatGPT, Goose, and VS Code.

In the context of knowledge management, MCP functions as a protocol layer connecting PKM tools (Obsidian, Notion, Logseq, etc.) with LLM agents. Details are covered in the next section. For MCP integration in the design tooling domain, see mcp-design-agent-integration.

6. PKM Tool Integration with LLMs

Three Architectural Patterns for Connectivity

PatternExampleCharacteristics
Plugin-embedded MCPObsidian Vault as MCPServer runs inside the app. Available only while the app is running
Standalone MCPmcp-logseq, reflect-mcpExternal process connecting to the app’s HTTP API
Hosted (SaaS)Notion MCP (mcp.notion.com/mcp)Vendor operates the server. OAuth authentication. Minimal setup

Obsidian

Obsidian has no first-party AI features. This is an intentional product decision: AI capabilities are delivered entirely through community plugins and users’ own API keys.

The two main plugins are Copilot for Obsidian (1.5M+ downloads, RAG chat over the entire vault, supporting OpenAI/Anthropic/Google/Ollama) and Smart Connections (semantic related-note display via local vector embeddings, no cloud required).

MCP connectivity is available through Vault as MCP (plugin-based, 13 operations, port 8765) and MCPVault (standalone, supporting Claude Desktop/Code/ChatGPT). Data remains local.

Notion

Notion deeply integrates AI features into the product. It offers Q&A (natural language search across the entire workspace), AI Connectors (cross-searching Slack/Drive/GitHub, etc.), Agents (introduced with Notion 3.0; Custom Agent reached GA in May 2026), and model selection (GPT-5.2/Claude Opus 4.5/Gemini 33).

The Notion MCP Server provides an official hosted endpoint (mcp.notion.com/mcp), connectable with a single command: claude mcp add --transport http notion https://mcp.notion.com/mcp. It features 22 tools and communication via Notion-Flavored Markdown (improving token efficiency). OAuth is required, which imposes constraints on fully automated workflows.

Logseq

Multiple MCP servers are under active community-driven development. mcp-logseq (v1.8.0, 16 core tools, semantic vector search, 300 stars) is the most mature.

Tana

Tana released its API and MCP server in early 2026. Supertags (a feature that attaches typed schemas to nodes) enable the most structured capture of any PKM tool. It includes an AI meeting agent (transcription in 60 languages with action item extraction).

Apple

Apple provides Writing Tools, recording summaries, and Image Wand for Notes. At WWDC 2026, it open-sourced the Foundation Models framework, opening a path for third parties to build AI apps that integrate with Notes. Notes MCP exists only as a community implementation (via AppleScript).

Google NotebookLM

NotebookLM suppresses hallucination through closed RAG, grounded exclusively in user-provided sources. A June 2026 update added Gemini 3.5, code execution notebooks, and agentic research features, pushing it from a note-taking tool toward a research execution layer. However, it is not positioned for long-term knowledge base construction.

7. GraphRAG: Fusing Knowledge Graphs with LLMs

Microsoft Research’s GraphRAG serves as the reference architecture. The pipeline consists of (1) extracting entities and relations from text, (2) constructing community hierarchies via Leiden clustering, (3) generating LLM summaries for each community, and (4) dispatching queries to Global Search (whole-corpus reasoning), Local Search (entity-neighborhood exploration), or DRIFT Search. LazyGraphRAG reduces indexing costs by 10-90%.

The 2026 research frontier lies in graph-based agent memory: a shift from passive logging to topological experience models. Graph extensions from Zep and Mem0 maintain temporal knowledge graphs as the memory layer for LLM agents.

The relationship with the Karpathy Wiki is complementary. GraphRAG “automatically extracts structure from unstructured text,” while the Karpathy Wiki has “the LLM directly maintain structured Markdown.” The former suits automatic indexing of large corpora; the latter suits collaborative knowledge management between humans and LLMs.

8. Fabric: Pattern-Based AI Pipelines

Daniel Miessler’s Fabric (v1.4.378, 2026-01-14) is a framework that decomposes AI use cases into reusable Patterns (structured system prompts) and composes them via CLI pipelines. It includes 300+ community-contributed patterns (summarize, extract_wisdom, explain_code, etc.) and supports arbitrary providers including OpenAI, Anthropic, and Ollama.

Miessler’s Personal AI Infrastructure (PAI) gives agents domain knowledge through a Skills directory (SKILL.md + workflows + tools), combining Fabric Patterns and MCP in the UNIX philosophy. Its center of gravity lies in input-transform-output pipeline processing rather than persistent knowledge accumulation.

9. RAG vs. Long Context: The 2026 Reality

The claim that “RAG is dead” recurs on social media, but enterprise RAG adoption reportedly grew 280% in 2025 [citation verification needed].

The prevailing winning pattern is a hybrid configuration: vector search narrows the evidence set, and a long-context model performs reasoning over it. When the corpus is small and stable (under 100K tokens), long context alone is advantageous; when the corpus is large and frequently updated, RAG is essential. The more accurate characterization is that “what is dead is not RAG but naive top-k vector search (lazy retrieval).”

Cross-Cutting Comparison: Pattern Classification

PatternKnowledge OwnerPersistenceDegree of StructurePrimary Use Case
Karpathy LLM Wiki / OKFLLM owns the wikiMarkdown files (git)High (index + wikilink + schema)Individual/team knowledge accumulation
CLAUDE.md / AGENTS.mdHuman (+ LLM auto memory)Files (git)Medium (freeform MD)Coding instructions and learning
Cursor RulesHuman.mdc filesMedium (frontmatter + glob)Coding conventions
Cline Memory BankHuman-defined, LLM-maintained6 MD filesHigh (fixed schema)Project context persistence
Fabric PatternsCommunityPattern filesMedium (templates)AI operation reuse
GraphRAGAuto-extractedGraph DBHigh (entities + relations)Reasoning over large corpora
PKM + MCPHuman (existing notes)Within PKM toolTool-dependentExposing existing knowledge to LLMs
NotebookLMHuman (provided sources)CloudMedium (closed RAG)Document comprehension and synthesis

Structural Observations

Three structures emerge.

(1) The subject of knowledge management has shifted. Traditional knowledge management asked “how should humans organize information.” LLM Wiki, OKF, and Context Engineering ask “how should knowledge be delivered to LLMs.” The agent of ownership and maintenance has not transferred from humans to LLMs; rather, LLMs have joined as consumers, and knowledge representation formats have begun to optimize for that consumer.

(2) Convergence on Markdown + filesystem + git. In-repo knowledge files (CLAUDE.md / AGENTS.md / Cursor Rules), the Karpathy Wiki, OKF, and Cline Memory Bank all build on Markdown + filesystem + git. Vector databases and graph databases remain in auxiliary roles. This aligns with the ETH Zurich finding that “human curation outperforms LLM auto-generation.” Transparency (human-readable and -writable), version control, and diff-ability are valued over the query performance of structured databases.

(3) MCP is redefining the value proposition of PKM tools. Obsidian and Notion were previously “knowledge management tools for humans.” Through MCP connectivity, they have begun to function also as “context sources for LLM agents.” However, maturity of integration varies. Notion is the most advanced with its official hosted MCP, Obsidian depends on community plugins, and Logseq and Tana remain at earlier stages of development.

References

← All Notes · Home