openclaw - 💡(How to fix) Fix ShieldCortex preamble advertises `remember`/`recall`/`forget`/`get_context` MCP tools that openclaw doesn't expose [1 participants]

Summary

The shieldcortex skill bundled with OpenClaw (v4.10.4 in this install) wires a SessionStart hook that instructs the model to proactively call MCP tools named remember, recall, forget, and get_context. The OpenClaw MCP server this release ships only exposes mcp__openclaw__memory_search and mcp__openclaw__memory_get — both read-only and file-backed. There is no write path from the model into the ShieldCortex SQLite store via MCP.

Result: the skill's contract and the host's MCP surface are out of sync. Models follow the skill's guidance, call tools that don't exist, and the user perceives the agent as unable to remember.

Evidence

• shieldcortex --version → 4.10.4
• /home/ubuntu/.openclaw/workspace/skills/shieldcortex/_meta.json → version 4.10.4, slug shieldcortex
• Deferred tool list exposed to the model (from <system-reminder> at session start): only mcp__openclaw__memory_search and mcp__openclaw__memory_get under the memory namespace. No remember/recall/forget/get_context anywhere.
• File-based memory (what memory_search/memory_get read) lives at MEMORY.md + memory/*.md per global CLAUDE.md — separate from the SQLite store at ~/.shieldcortex/memories.db.

What I'd expect from OpenClaw

One of:

1. Ship the write-side MCP tools. Add mcp__openclaw__memory_remember / memory_forget that write to ~/.shieldcortex/memories.db, so the skill's preamble is honest.
2. Declare a skill/MCP compatibility contract. Skills should declare the MCP tools they require; OpenClaw should warn at skill load time if the host's MCP server doesn't expose them (or refuse to wire the hook).
3. Document the split. If the intent is "file-based memory (MEMORY.md) is the source of truth and ShieldCortex is advisory", the global CLAUDE.md and the skill preamble both need to say so — right now they contradict each other.

Related

Filed against ShieldCortex separately (bug: SessionStart hook fires on every resume, plus tool-name mismatch). This issue is about the host-side contract gap, not the skill's internal behavior.