One shared memory for every coding agent: prove the configs actually point at the same server

The Stack

Steps

1. 1

   Scaffold each agent's MCP config + a git vault, then assert one shared brain

   Write the same memory server into three agents' configs, put a write-the-memory policy note in the vault, and git-init the vault. The check fails loudly if any agent points at a different server (memory would not be shared), if the write policy is missing (MCP will not auto-journal), or if the vault is not under git (write access with no undo). Change one server command and the identity check trips, which is the whole point.

   rm -rf vault .cursor .mcp.json cline_mcp_settings.json
   mkdir -p vault/memory .cursor
   SERVER='{ "mcpServers": { "basic-memory": { "command": "uvx", "args": ["basic-memory", "mcp"] } } }'
   printf '%s' "$SERVER" > .mcp.json
   printf '%s' "$SERVER" > .cursor/mcp.json
   printf '%s' "$SERVER" > cline_mcp_settings.json
   cat > vault/memory/policy.md <<'MD'
   # House rule: write what we decide to memory
   When we make a decision or learn a project fact, save a short note here so the
   next agent inherits it. Do not assume "it is in memory" means "it is current".
   MD
   git -C vault init -q && git -C vault add -A && git -C vault -c user.email=a@b.c -c user.name=x commit -qm init
   
   git -C vault rev-parse --is-inside-work-tree >/dev/null 2>&1 || { echo "BAD: vault is not a git work tree; grant write access only to a version-controlled store"; exit 1; }
   
   node -e '
   const fs = require("fs");
   const agents = { "Claude Code": ".mcp.json", "Cursor": ".cursor/mcp.json", "Cline": "cline_mcp_settings.json" };
   const SERVER = "basic-memory";
   function spec(s) { if (!s) return null; if (s.url) return "url:" + s.url; return "cmd:" + s.command + " " + ((s.args || []).join(" ")); }
   const specs = [];
   for (const name of Object.keys(agents)) {
     const srv = (JSON.parse(fs.readFileSync(agents[name], "utf8")).mcpServers || {})[SERVER];
     if (!srv) { console.log("BAD: " + name + " has no " + SERVER + " MCP server configured"); process.exit(1); }
     specs.push(spec(srv));
   }
   const uniq = Array.from(new Set(specs));
   if (uniq.length !== 1) { console.log("BAD: agents resolve to DIFFERENT servers, memory is not shared: " + JSON.stringify(specs)); process.exit(1); }
   if (!fs.existsSync("vault/memory/policy.md")) { console.log("BAD: no memory-write house rule; MCP does not auto-journal"); process.exit(1); }
   console.log("shared-memory OK: 3 agents (Claude Code, Cursor, Cline) all resolve to one server [" + uniq[0] + "]; write-the-memory policy present; vault is git-tracked. The model deciding WHEN to save is fenced");
   '

2. 2

   Point the agents at the server and run for real (the model steps, not checked by CI)

   Install the server (uv tool install basic-memory, or run an Obsidian MCP server over your existing vault), then add the validated config to each agent (Claude Code /ide or .mcp.json, Cursor and Cline MCP settings; Codex uses the same server via its config.toml). Ask an agent to save a decision, then open a different agent and confirm it reads the same note. Whether the agent actually journals what matters, and whether a recalled note is still current, are the fenced parts, no green check can promise either.