One Folder Runs Claude, Codex, Gemini — and Obsidian. The Anti-Lock-In Memory Pattern Going Around

A recent walkthrough from the ICOR with Tom channel (Tom Solid) makes a pointed argument to anyone who lives in an AI CLI all day: stop renting your memory from a vendor. The video, "One Folder Runs Claude, Gemini, Codex, and even Obsidian (Free on GitHub)," demos a free, source-available scaffold called myPKA — a plain-Markdown folder that any agent CLI can adopt as its brain, then hand off to a different CLI tomorrow without losing a thing.

For the vybecoding.ai audience the interesting part is not the personal-knowledge angle. It is the architecture: a portable convention that decouples your accumulated context, operating rules, and agent roster from whichever model you happen to be paying for this month.

The Core Argument: Turn Off Vendor Auto-Memory

The video opens with a deliberately contrarian claim — switch off the built-in "memory" feature in Claude, ChatGPT/Codex, Gemini, or whatever you use. The reasoning is twofold. First, auto-memory grabs context opportunistically; it decides what is "useful" about you, and you cannot easily inspect or correct it. Second, and more important for engineers, that memory is trapped. The moment you switch CLIs — because one is cheaper, faster, not rate-limited that hour, or because security policy demands a local model — your accumulated context does not come with you. You are back to a cold start.

The proposed fix is a single sentence with real architectural weight: the folder, not the agent, is the source of truth. If durable knowledge lives in plain Markdown inside a directory, memory becomes a property of the workspace. Every agent that can read files inherits it, and switching models becomes a routing decision instead of a memory wipe.

The Adapter-Prompt Trick

The mechanism that makes this tool-agnostic is one file at the repo root: ADAPTER-PROMPT.md. You open the folder in any LLM tool and tell it to "initialize yourself in this folder" (or run the tool's native init). The LLM reads the adapter prompt and writes a thin, tool-specific pointer file — CLAUDE.md for Claude Code, GEMINI.md for Gemini CLI, AGENTS.md (which Codex and Cursor read natively, so Codex needs almost no setup).

The design intent is the opposite of the prevailing "install more skills and plugins" trend. The pointer files are kept deliberately slim and simply reference the agent definition files instead of bundling vendor-specific skill packs. The argument: pre-made skill bundles load more tokens than you need and quietly re-introduce lock-in. In the demo, three CLIs — Claude (Cowork and terminal), Codex, and Gemini — are pointed at copies of the same folder and each initializes into the identical behavior with no per-tool prompt engineering.

Memory You Own, Not Memory You Rent

With vendor auto-memory off, persistence comes from session logs. Saying "close the session" writes a Markdown log into the scaffold's Team Knowledge area. Because logs are plain files, they are reviewable and time-ordered: you can tell the agent "last week worked better — review the logs" and it can actually do that. The latest version of the scaffold (v1.10.0) adds per-agent journals so individual specialists accumulate durable cross-session insight rather than dumping everything into one file. This is the same idea behind context engineering: explicit, inspectable, version-controllable memory beats an opaque vendor toggle.

The Agent Team

myPKA ships with a nine-specialist roster defined entirely in Markdown — no proprietary agent format. The three the video foregrounds are the load-bearing ones: Larry, the orchestrator and single point of contact, who routes every request via an agent index; Nolan, the talent-acquisition agent, who handles hiring a new specialist end-to-end when the team needs a capability it lacks; and Pax, the deep-research agent, who produces triangulated briefs. The full roster also includes Penn (journal writer), Mack (automation and MCP/API integration), Silas (database architect), Charta (infographics), Pixel (visual generation), and Iris (design-system architect). The pattern — an orchestrator that hires and dispatches sub-agents whose contracts live in version-controlled Markdown — is one developers can lift wholesale, independent of the personal-knowledge use case.

Obsidian Out of the Box, SQLite as a One-Prompt Upgrade

Because every file is Markdown with [[wiki links]] and YAML properties, Obsidian opens the same folder with zero setup — knowledge graph, backlinks, and embeds all work without an Obsidian-specific CLI or required plugin. That is the "even Obsidian" in the title: it is not an integration, it is a side effect of writing portable Markdown. For teams that outgrow flat files, a single prompt ("switch to SQLite") converts the store into a SQLite database for more efficient retrieval at scale, while Markdown remains the recommended default for most users. Multi-device sync via Dropbox, iCloud, or Drive works fine sequentially; the only failure mode is editing the same folder on two devices in parallel.

Why Developers Should Care

Strip away the personal-knowledge framing and what is left is a reusable answer to a real problem: agent memory and behavior should be a property of the repo, not the vendor. The ADAPTER-PROMPT.md → per-tool pointer-file convention is trivial to adopt in any project, and the orchestrator/hiring/researcher split is a clean multi-agent topology. If your team rotates between Claude Code, Codex, and Gemini for cost or rate-limit reasons — increasingly common in 2026 — a folder-first memory layer turns those switches from painful resets into routine routing.

The Caveats

Be precise about the license. myPKA is published under CC BY-NC-SA 4.0 — free for personal, non-commercial use with attribution and share-alike. It is "free and source-available," not open source in the OSI sense, and the non-commercial clause matters if you want to build it into a commercial product. "ICOR" and "PAPERLESS MOVEMENT" are registered trademarks. There is also a paid course that teaches the build step by step; the scaffold itself is the free GitHub download and stands on its own. None of the productivity claims need the course to be useful — the architecture is the takeaway.

The Bottom Line

This is not a new model or a benchmark. It is a convention — and a good one. The lasting idea is that a plain-Markdown folder plus a vendor-neutral adapter prompt gives you portable, inspectable agent memory that survives a CLI switch. That is a developer-experience pattern worth copying even if you never touch the personal-knowledge layer it ships with.

_Published on May 16, 2026 by vybecoding.ai editorial team_