claude-mem

name	claude-mem
description	Search the claude-mem persistent database for past context. Tool use is auto-captured — use this skill to search and recall. Standing rules and preferences go in MEMORY.md instead.
allowed-tools	Bash(curl:*)

Persistent Memory

You have a persistent memory database that survives across sessions. Requires $CLAUDE_MEM_URL environment variable. If not configured, run /add-claude-mem on the host to set it up.

Important: Always use project=nanotars in all API calls. The hook writes observations under this project for every group, so queries scoped to it surface the full history.

What's Captured Automatically

Every tool use (Bash, Read, WebSearch, MCP calls, etc.) is automatically saved to the database with the tool name, input, and output. You don't need to manually save tool results — they're already searchable. Relevant prior context is also auto-injected into the system prompt at session start, so you start each turn with recent memory in view; only reach for explicit search when the auto-injected slice doesn't cover what you need.

Where to Store What

What	Where	Why
Standing rules ("always use metric")	MEMORY.md	Auto-loaded every session, always visible
Personal facts ("user likes flat whites")	MEMORY.md	Should be available without searching
Preferences and routines	MEMORY.md	Persistent across all conversations
Important conclusions from this conversation	claude-mem save	Searchable later, not needed every session
"Remember this for later" (user request)	claude-mem save	Explicit user request to persist a fact
Tool outputs and actions taken	Auto-captured	Already saved, no action needed

Manual Save

Use the save endpoint for important conclusions or facts the user shares that aren't from a tool call:

curl -s -X POST "$CLAUDE_MEM_URL/api/memory/save" \
  -H "Content-Type: application/json" \
  -d '{"text": "User prefers flat white coffee", "project": "nanotars"}'

When to Search Memory

• User asks about something discussed previously
• User references a person, project, or recurring topic
• User says "remember when...", "last time...", or "did I tell you..."
• You need context about past decisions or plans
• Before making assumptions about recurring topics
• Be proactive — search memory at the start of conversations about recurring topics

Search Memory

curl -s "$CLAUDE_MEM_URL/api/search?query=morning+routine+preferences&project=nanotars" | jq -r '.content[0].text // .'

Search returns an index with observation IDs and titles. If you need full details for specific results, fetch them by ID.

Get Full Details

curl -s -X POST "$CLAUDE_MEM_URL/api/observations/batch" \
  -H "Content-Type: application/json" \
  -d '{"ids": [42, 43]}' | jq '.[].narrative // .[].text'

Get Timeline Context

See what happened around a specific observation:

curl -s "$CLAUDE_MEM_URL/api/timeline?anchor=42&project=nanotars"

Tips

• Use broad search queries to find related memories (e.g., "coffee preferences" not just "flat white")
• Use timeline to understand the context around a specific observation
• MEMORY.md = things you need every session (rules, preferences, personal facts)
• claude-mem save = things you might need later (decisions, conclusions, user requests to remember)
• Don't manually save tool outputs — they're already auto-captured