memory-search

name	memory-search
description	Search past Claude chat history using the memory-bank vector DB. Use when the user asks about previous conversations, past solutions, or wants to find something discussed in an earlier session.
allowed-tools	Bash, Read

Memory Bank Search Skill

Search semantically over ingested Claude Code and Claude Desktop chat histories stored in a local Qdrant vector DB.

When to use

Trigger this skill when the user asks things like:

• "Did we talk about X before?"
• "What was that solution we used for Y?"
• "Find past conversations about Z"
• "What have I worked on related to X?"
• "Search my chat history for..."

Prerequisites

Before searching, verify the tool is installed:

memory-bank stats

If the command is not found, the user needs to install memory-bank:

uv tool install memory-bank
memory-bank setup install

If the DB is empty (0 messages), ingest first:

memory-bank ingest claude-code

Search — use --agent mode

Always use --agent when calling from Claude Code. It returns compact JSON (no id field, date-only timestamps, 300-char content snippets, score ≥ 0.5 filter) that costs ~60% fewer tokens than --json while preserving all signal needed to answer the user.

memory-bank search "your query" --agent

In --agent mode, min-score defaults to 0.5 and limit defaults to 5. Override when needed:

# More results, lower score bar, longer snippets
memory-bank search "query" --agent --limit 10 --min-score 0.3 --snippet 500

# Scope to a project or role
memory-bank search "query" --agent --project my-app --role assistant

# Filter by source
memory-bank search "query" --agent --source claude-code

# Filter by message category (bugfix, feature, refactor, decision, research)
memory-bank search "authentication" --agent --category bugfix

# Include surrounding context to understand whether a result was a solution or dead-end
memory-bank search "docker fix" --agent --context 2

# Remove duplicate results across sessions (keeps highest-scoring copy)
memory-bank search "docker fix" --agent --dedupe

# Scope to the current working directory's project
memory-bank search "refactor" --agent --current-project --dedupe

Human-readable table (for showing results to the user)

memory-bank search "query"

Full JSON (for scripting / jq pipelines)

memory-bank search "query" --json --snippet 400

Environment variable for custom DB path

MEMORY_BANK_DB=/custom/path memory-bank search "query" --agent

Workflow

1. Run memory-bank search "..." --agent
2. Parse the compact JSON array: each object has score, role, src, date, text, and optionally proj / sid
3. Summarize the most relevant findings to the user, quoting briefly
4. If results are empty or low quality:
   • Rephrase the query more semantically (e.g. "container networking" instead of "docker errors")
   • Lower --min-score to 0.3
   • Remove or broaden --since / --before filters
   • Increase --limit to 20

Ingest commands reference

memory-bank ingest claude-code                         # auto-detects ~/.claude/projects/
memory-bank ingest claude-code --path /custom/path
memory-bank ingest claude-desktop --path ~/Downloads/conversations.json
memory-bank ingest all

Other commands

memory-bank stats
memory-bank ui           # web UI at http://localhost:8765
memory-bank hooks install    # auto-ingest after each session
memory-bank hooks status

Web UI

memory-bank ui           # opens http://localhost:8765
memory-bank ui --port 8765

Auto-ingest hooks

memory-bank hooks install    # run ingest automatically after each session
memory-bank hooks status