Memory Manager

Memory Manager

Complete memory management system for agents, including semantic search, file management, and memory curation workflows.

When to Use This Skill

• Searching for past decisions, preferences, or context
• Reading or writing to memory files
• Managing daily logs and long-term memory
• Organizing and archiving old memory files
• Before answering questions about prior work or conversations

Core Memory Tools

에이전트에 내장된 두 개의 메모리 도구 (langgraph_agent.py에 @tool로 구현):

memory_search(query, max_results?, min_score?)

MEMORY.md + memory/*.md 파일에 대한 키워드 기반 검색.

Use when:

• User asks about past events, decisions, or preferences
• Looking for related information even with different wording
• Need to recall context from previous sessions

Parameters:

• query (required): Search query string
• max_results (optional, default: 5): Max results to return
• min_score (optional, default: 0.0): Minimum relevance threshold (0.0-1.0)

Returns: JSON array of snippets with text, path, from (line), lines, score.

이 도구는 에이전트가 직접 호출합니다 (execute_code 불필요).

memory_get(path, from_line?, lines?)

특정 메모리 Markdown 파일을 직접 읽기.

Use after:

• memory_search to get full context
• When you know the exact file path

Parameters:

• path (required): Workspace-relative path (e.g., "MEMORY.md", "memory/2026-02-27.md")
• from_line (optional, default: 0): Starting line number, 1-indexed (0 = read from beginning)
• lines (optional, default: 0): Number of lines to read (0 = read entire file)

Returns: JSON with text (file content) and path.

Graceful degradation: If file doesn't exist, returns { "text": "", "path": "..." } (no error).

Memory File Structure

MEMORY.md (Long-term memory)

• Curated, important information
• Decisions, preferences, durable facts
• Security: Only loaded in main, private session (not group chats)

memory/YYYY-MM-DD.md (Daily logs)

• Day-to-day notes, running context
• Append-only during the day
• Today + yesterday loaded at session start

Workflow Examples

에이전트는 이 도구들을 직접 호출합니다 (execute_code 불필요):

1. Search then read detailed context

1. memory_search(query="Tavily API setup") 호출
2. 결과에서 가장 관련성 높은 항목의 path, from, lines 확인
3. memory_get(path=결과.path, from_line=결과.from, lines=결과.lines) 호출

2. Check today's notes

1. memory_get(path="memory/2026-03-02.md") 호출
2. text가 비어있으면 아직 오늘의 로그가 없음

3. Search across time

1. memory_search(query="Gmail configuration", max_results=5) 호출
2. 결과의 path에서 날짜 확인 (e.g., "memory/2026-02-27.md")
3. memory_get(path="memory/2026-02-27.md") 로 전체 내용 확인

File Management Utilities

Use scripts/manage_memory.py for file operations:

Create daily log

# Create today's log
python scripts/manage_memory.py create-daily

# Create specific date
python scripts/manage_memory.py create-daily --date 2026-03-01

Append content

# Append to MEMORY.md
python scripts/manage_memory.py append MEMORY.md "New important fact"

# Append to daily log with section
python scripts/manage_memory.py append memory/2026-03-01.md \
  "Meeting notes here" --section "Meetings"

List recent logs

# List last 7 days
python scripts/manage_memory.py list

# List last 30 days as JSON
python scripts/manage_memory.py list --days 30 --json

Archive old logs

# Archive logs older than 90 days
python scripts/manage_memory.py archive --days 90

Best Practices

When to Write Memory

1. MEMORY.md - Durable, important facts:

   • User preferences and settings
   • Important decisions and their reasoning
   • API keys and credentials (redacted if sensitive)
   • System configurations
   • Long-term project information

2. memory/YYYY-MM-DD.md - Daily context:

   • What happened today
   • Tasks completed
   • Meetings and conversations
   • Temporary notes and observations
   • Links to resources used

3. When someone says "remember this" - Write it down immediately!

   • Don't keep "mental notes" - they vanish on session restart
   • Memory files are the ONLY persistence

Search Before Answering

MANDATORY: Before answering questions about:

• Prior work or decisions
• Past conversations
• User preferences
• Dates and timelines
• People and relationships
• TODOs and tasks

Always run memory_search first, even if you think you remember. The current session context may not include relevant past information.

Curation Workflow

Periodically (during heartbeats or when memory is full):

1. Read recent memory/YYYY-MM-DD.md files
2. Identify important facts worth keeping long-term
3. Update MEMORY.md with distilled learnings
4. Remove outdated info from MEMORY.md
5. Archive old daily logs

Think: Daily files = raw notes, MEMORY.md = curated wisdom.

Security Considerations

• MEMORY.md only loads in main session (direct chat with user)
• Never load in group chats to prevent information leakage
• Redact sensitive information (passwords, tokens) before writing
• User can always read the files directly - treat them as shared knowledge

Advanced: Memory Search Configuration

Memory search uses vector embeddings for semantic search. Common configurations:

Hybrid Search (BM25 + Vector)

Best for:

• Finding exact IDs or code symbols
• Semantic queries with different wording

MMR Re-ranking

Enable when you see redundant results:

• Balances relevance with diversity
• Prevents multiple similar snippets

Temporal Decay

Enable for long-running agents:

• Recent memories rank higher
• Old information naturally fades

For detailed configuration, see references/memory-system.md.

Common Patterns

Daily standup / summary

const yesterday = new Date(Date.now() - 86400000).toISOString().split('T')[0];
const yesterdayLog = await memory_get(`memory/${yesterday}.md`);

// Summarize what happened yesterday
// Write today's plan to today's log

Project context recall

// Search for project information
const projectInfo = await memory_search("project X status", 3);

// Get full context from most relevant result
const context = await memory_get(projectInfo[0].path);

Preference lookup

// Check user preferences
const prefs = await memory_search("preferred email client", 2);

// Fall back to asking if not found
if (prefs.length === 0 || prefs[0].score < 0.7) {
  // Ask user for preference
}

Troubleshooting

No search results

• Check if memory files exist (memory_get the file directly)
• Verify embedding provider is configured
• Try different query wording

Search too slow

• Enable hybrid search
• Use remote embeddings instead of local
• Reduce candidateMultiplier in config

Redundant results

• Enable MMR re-ranking (mmr.enabled = true)
• Increase diversity (lower lambda)

Stale information ranking high

• Enable temporal decay (temporalDecay.enabled = true)
• Adjust halfLifeDays (lower = faster decay)

Reference Documentation

For complete technical details, see references/memory-system.md:

• Full tool specifications
• Configuration options
• Vector search backends
• QMD experimental backend
• Session memory indexing
• Troubleshooting guide

Example: Full Memory Workflow

// 1. User asks: "What did we decide about Gmail setup?"

// 2. Search memory
const results = await memory_search("Gmail setup decision", 3);

// 3. Get detailed context
let context = "";
for (const result of results) {
  const detail = await memory_get(result.path, result.from, result.lines);
  context += `\n--- ${result.path} ---\n${detail.text}\n`;
}

// 4. Answer based on retrieved context
// "Based on our conversation on 2026-02-27, we decided to..."

// 5. If new decision made, write it to today's log
const today = new Date().toISOString().split('T')[0];
await memory_get(`memory/${today}.md`); // Ensure exists
// Then use file tools to append the new decision

Notes

• Memory files are plain Markdown - you can read/write them directly
• Changes to memory files trigger reindexing (debounced)
• memory_search and memory_get는 langgraph_agent.py에 @tool로 구현되어 있음
• 에이전트가 직접 도구로 호출 가능 (execute_code를 통한 호출 불필요)
• This skill provides management utilities and usage patterns
• Memory is per-agent - each agent has its own workspace and memory index