memory-workflow

AgentKits Memory Workflow

When to Activate

Use this skill when:

• User asks about past work, previous sessions, or what was done before
• User references a decision, pattern, or error you don't have context for
• You need project history, conventions, or architectural decisions
• User asks "what did we do about X?" or "how did we handle Y?"
• You're missing context that should exist from earlier sessions
• Starting work on a feature that may have prior decisions recorded

Prerequisites

Before searching, check if memories exist:

memory_status()

If the database is empty, skip recall and inform the user.

3-Layer Search Workflow

Layer 1: Search Index (lightweight, ~50 tokens/result)

memory_search(query="your search term")

• Returns IDs, titles, categories, dates, and relevance scores
• Filter by category: decision, pattern, error, context, observation
• Filter by date: dateStart="2025-01-01", dateEnd="2025-12-31"
• Sort: orderBy="relevance" (default), "date_asc", "date_desc"

Layer 2: Timeline Context (understand what happened around a result)

memory_timeline(anchor="MEMORY_ID")

• Shows what happened before/after a specific memory
• Helps understand the sequence of events
• Use when you need temporal context

Layer 3: Full Details (only for filtered IDs)

memory_details(ids=["ID1", "ID2"])

• Returns complete content for selected memories
• Limit to 3-5 IDs at a time to conserve tokens
• NEVER fetch details without filtering through Layer 1 first

Quick Topic Recall

For a fast overview of everything known about a topic:

memory_recall(topic="authentication")

This returns a grouped summary. Follow up with memory_details for specifics.

Saving Memories

Save important information for future sessions:

memory_save(content="...", category="decision", tags="auth,security", importance="high")

Categories: decision, pattern, error, context, observation Importance: low, medium, high, critical

Token Efficiency Rules

1. ALWAYS start with memory_search (Layer 1), never jump to memory_details
2. Review search results and select only relevant IDs before fetching details
3. Use filters (category, date range) to narrow results
4. Limit memory_details to 3-5 IDs per call
5. This workflow saves ~87% tokens vs fetching everything at once