memory-search-bridge

name	memory-search-bridge
description	Search persistent memory across mempalace MCP, claude-mem MCP, and the structured filesystem fallback. Use when the user asks "how did we solve X before?", "what did we decide about Y?", or anything that implies recalling prior-conversation context.

Memory Search Bridge

Use this skill when the user references prior decisions, prior fixes, or anything that should already be persisted.

Step 1 — Determine the active backend

Read ~/.claude/sustain/state.json and check detection.preferred. The plugin's SessionStart hook refreshes this every session.

Step 2 — Route the search

mempalace preferred

Use mempalace's MCP search tools. Common entry points:

• mcp__mempalace__search — query across all wings.
• mcp__mempalace__list_wings — get a wing inventory if the user asked about a category rather than a fact.
• mcp__mempalace__get_drawer — fetch a specific entry by id when you know it.

Mempalace stores verbatim entries — show the matching content directly; no need to summarise unless the user asks.

claude-mem preferred

Use claude-mem's MCP search tools (namespace mcp__plugin_claude-mem_mcp-search__):

• smart_search — best general-purpose entry point.
• search — keyword search.
• smart_outline / smart_unfold — when the user wants structure or to drill down.
• timeline — when the user is asking "when did we…".
• get_observations — fetch raw observations by id.

claude-mem returns compressed observations; if a result feels paraphrased and the user needs verbatim, fall back to fs (which mirrors recent writes) before grepping the codebase.

fs preferred (or as fallback after either of the above)

Search ~/.claude/sustain/memory/:

grep -rli "<keyword>" ~/.claude/sustain/memory/wings/ | head -20

or use the plugin's adapter:

import { query } from "<plugin-root>/lib/memory/adapter.js";
query({ text: "your query", wing: "feedback" /* optional */ });

The adapter returns hits with excerpts and (when a richer backend is preferred but the call landed here) a routing suggestion you should also try.

Step 3 — If nothing found

Before searching the codebase from scratch, consider:

• Was the original question answered in the current conversation? If the user says "earlier" they may mean earlier in this session — re-read recent turns first (R5.1).
• Could it be in ~/.claude/projects/<project>/memory/MEMORY.md? That's the user's auto-memory, separate from the plugin's store.
• Could it be in the user's ~/Documents/AI/ notes? Sometimes durable knowledge lives in long-form markdown rather than memory entries.

Only after these turn up empty, search the codebase or ask the user to refresh your memory.

Step 4 — Trust but verify

Memory backends can return stale entries. If a recalled fact contradicts current code or git state, trust what you observe now and update or delete the stale memory rather than acting on it (R3.5 logic applied to memory).