Agent Memory: The Format Is Markdown — But the Problem Isn't

Three independent actors — a former OpenAI research director, a Google engineering team, and a prominent startup investor — converged on the same format for agent memory in 2026. The format is Markdown. The more useful part is that they’re solving different problems with it.

What Happened

Writing in The New Stack, Janakiram MSV documented the Markdown convergence across three notable practitioners:

Andrej Karpathy published a GitHub Gist describing his “LLM Wiki” — a personal knowledge system built from interlinked Markdown files designed to function as a portable, AI-accessible memory layer. The design is explicit: these files travel with Karpathy across AI tools and sessions, serving as persistent context that no single product owns or locks down.

Google shipped what they’re calling the Open Knowledge Format — a proposed standard for enterprise-scale agent memory using structured Markdown as the interchange layer. The goal is cross-tool portability and a common substrate for agents operating across enterprise systems and vendors.

Garry Tan (Y Combinator) independently articulated a similar pattern for startup use cases: Markdown files as durable artifacts of decisions, reasoning, and operational context — designed to outlast any specific tool or AI product.

The convergence wasn’t coordinated. Each arrived at Markdown independently, for different reasons, targeting different audiences.

Why It Matters

The surface-level takeaway — “use Markdown for AI context files” — misses the more important signal.

Karpathy is solving a portability and ownership problem: AI tools come and go, but a Markdown file lives on the filesystem and travels with you. The format is the persistence layer, not the application.

Google is solving an enterprise interoperability problem: agents across departments need to share context in a format that isn’t locked to one vendor’s schema or API surface.

Tan is solving a knowledge-decay problem: startup institutional memory disappears when people leave or tools change, and durable text files resist both failure modes better than any platform-hosted note system.

These are three different architectural bets. They all use Markdown because it’s human-readable, processable by LLMs, and portable across systems — but the reason it matters differs by context.

The line from MSV’s piece that cuts deepest: “the real moat is shifting from models to files.” The agent model is commoditizing. What differentiates outcomes is the quality and structure of the context the agent has access to — and that context increasingly lives in plain text.

What This Means for Builders

Pick the problem before picking the pattern. If you’re building personal AI tooling, Karpathy’s portable-file-as-memory approach is the right frame. If you’re building enterprise multi-agent systems, Google’s interoperability-first framing matters more. Conflating them produces architectures optimized for neither.

Treat your Markdown files as first-class artifacts. Version them in git, review changes when they update, and document the reasoning alongside the conclusion. A context file that captures “we do X because Y” is more durable than one that only captures “we do X.”

Start before you’re ready. The best time to build your LLM Wiki (or enterprise equivalent) is before you have a critical number of agents in production. The organization’s invisible knowledge layer — the institutional judgment that never made it into any system (see: The Organizational Iceberg) — is easier to capture while it’s still fresh than after it’s disappeared from institutional memory.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.