Practical guide
When a local
memory folder
is enough,
and when it isn't
Local shared-memory tools for coding agents are having a moment. This week one of them was trending on GitHub, and the pitch is good: quit Claude Code mid-task, open Codex in the same folder, and it picks up the handoff from plain markdown files.
If all your AI work happens in coding agents on one machine, in one repo, that is often enough. Use it.
Where a folder stops
It stops being enough in three situations:
The work leaves the IDE.
Planning in ChatGPT, writing in Claude on the web, reviewing on your phone. Those apps can't read a folder on your laptop. They can read a remote MCP server.
You switch machines.
A folder in one checkout doesn't follow you to the other laptop, or to a cloud agent that starts from a fresh clone.
The "why" gets lost in the noise.
Long memory files grow. A handoff that works carries five things: the goal, decisions and why, what's done, what's still unverified, and one next step.
Hosted project state
Arkon Vault is that small state, hosted.
ChatGPT, Claude, Codex and Cursor read and write it over MCP (https://vault.arkoncybersecurity.com/mcp), with OAuth so only you can see it. It sits next to a local memory folder rather than replacing it: keep your detailed notes in the repo, keep the cross-tool state in Vault.