Agent Memory Ultimate

Agent Memory Ultimate

Platform: OpenClaw
Docs: https://docs.openclaw.ai

Overview

AI agents wake fresh each session — no memory of yesterday. This skill solves that by implementing a memory system modeled on human cognition:

Why it works: Humans don't remember everything — they consolidate important patterns during sleep. This skill gives agents the same architecture.

File Structure

workspace/
├── MEMORY.md              # Long-term memory (curated)
├── AGENTS.md              # Operating instructions
├── SOUL.md                # Identity & personality
├── USER.md                # Human profile
├── db/
│   └── agent.db           # SQLite for structured data
├── bank/
│   ├── entities/          # People profiles
│   │   ├── PersonName.md
│   │   └── ...
│   ├── contacts.md        # Quick contact reference
│   └── opinions.md        # Preferences, beliefs
└── memory/
    ├── YYYY-MM-DD.md      # Daily logs
    ├── projects/
    │   ├── index.md       # Master project list
    │   └── project-name/
    │       ├── index.md   # Project overview
    │       └── sprints/   # Phase subfolders
    └── knowledge/         # Topic-based docs

For OpenClaw workspace setup, see the workspace docs.

Organizing by Person

Store entity profiles in bank/entities/:

# PersonName.md

## Basic Info
- **Name:** Full Name
- **Phone:** +1234567890
- **Relationship:** Friend / Family / Colleague

## Context
How you know them, key interactions, preferences

## Notes
Running log of important details learned

When to create an entity file:

• Recurring person in conversations
• Someone with specific preferences to remember
• Family members, close contacts

Organizing by Project

Projects live in memory/projects/<project-name>/:

memory/projects/
├── index.md                    # Master list of all projects
├── website-redesign/
│   ├── index.md                # Project purpose, status, decisions
│   ├── architecture.md         # Technical design
│   └── sprints/
│       └── 2026-02/
│           └── index.md        # Sprint goals, progress
└── home-automation/
    ├── index.md
    └── devices.md

Project index.md template:

# Project Name

**Status:** Active / Paused / Complete
**Started:** YYYY-MM-DD
**Purpose:** One-line summary

## Key Decisions
- [Date] Decision and rationale

## Links
- Repo: ...
- Docs: ...

SQLite for Structured Data

Use SQLite (db/agent.db) for data that needs queries. OpenClaw agents can execute SQL directly via the exec tool.

When to Use SQLite vs Markdown

Schema Example

-- Contacts table with full-text search
CREATE TABLE contacts (
  id INTEGER PRIMARY KEY,
  phone TEXT UNIQUE,
  name TEXT,
  notes TEXT,
  source TEXT,
  created_at TEXT DEFAULT CURRENT_TIMESTAMP
);

CREATE VIRTUAL TABLE contacts_fts USING fts5(
  name, phone, notes,
  content='contacts',
  content_rowid='id'
);

-- Conversation memory
CREATE TABLE messages (
  id INTEGER PRIMARY KEY,
  chat_id TEXT,
  sender TEXT,
  body TEXT,
  timestamp INTEGER,
  channel TEXT
);

CREATE VIRTUAL TABLE messages_fts USING fts5(
  body, sender,
  content='messages',
  content_rowid='id'
);

Query Examples

-- Find contact by partial name
SELECT * FROM contacts_fts WHERE contacts_fts MATCH 'john*';

-- Search conversation history
SELECT * FROM messages_fts WHERE messages_fts MATCH 'project AND deadline';

-- Recent messages from person
SELECT * FROM messages 
WHERE sender LIKE '%5551234%' 
ORDER BY timestamp DESC LIMIT 20;

Daily Cycle

1. Wake Up (Session Start)

Add this to your AGENTS.md:

Before doing anything:
1. Read SOUL.md — who you are
2. Read USER.md — who you're helping
3. Read memory/YYYY-MM-DD.md (today + yesterday) — recent context
4. If main session: Also read MEMORY.md

2. During Day (Active Session)

• Write significant events to memory/YYYY-MM-DD.md
• Don't rely on "mental notes" — they don't survive restarts
• When told to remember something: write it NOW

3. Sleep Cycle (Consolidation)

Schedule a daily "sleep" task using OpenClaw cron:

{
  "schedule": { "kind": "cron", "expr": "0 3 * * *", "tz": "Europe/Madrid" },
  "payload": { 
    "kind": "systemEvent", 
    "text": "Memory consolidation: Review recent daily logs, extract key learnings, update MEMORY.md, prune outdated info." 
  },
  "sessionTarget": "main"
}

What consolidation does:

1. Reviews last 3-7 daily logs
2. Extracts patterns and recurring lessons
3. Adds distilled insights to MEMORY.md
4. Removes outdated information

Memory Types

Raw Memory (Daily Logs)

• Everything that happened
• Decisions made and why
• Errors and lessons
• Technical details

Curated Memory (MEMORY.md)

• Abstract principles (not specific fixes)
• User preferences
• Core lessons that apply broadly
• Things that should survive months/years

Abstraction example:

• ❌ Daily log: "Fixed senderE164 bug in message-line.ts"
• ✅ MEMORY.md: "Chat ID ≠ Sender — always verify actual sender field"

Why This Works (Cognitive Science)

1. Spaced repetition — Daily review reinforces important memories
2. Active consolidation — Humans consolidate during sleep; agents consolidate in scheduled tasks
3. Chunking — MEMORY.md groups related concepts into retrievable chunks
4. Forgetting is useful — Raw logs fade in relevance; curated memory persists
5. External memory — Files ARE your memory (like notes for humans)

Memory Tiers Summary

Rule of thumb:

• Need to search across many records → SQLite
• Need to read/update narrative context → Markdown
• Need instant access → Session context (but dies on restart)

Scripts & CLI

This skill includes ready-to-use Python scripts for managing your knowledge base.

Quick Start

# Initialize database (first time)
python3 scripts/init_db.py

# Query your data
python3 scripts/query.py search "term"
python3 scripts/query.py contact "+1555..."
python3 scripts/query.py stats

Available Commands

Data Sources

WhatsApp Contacts & Groups

Export via the whatsapp-ultimate skill's contact extraction:

python3 scripts/sync_whatsapp.py

Indexes: contacts, groups, memberships (who is in which group).

ChatGPT Conversation History

Export from ChatGPT and place in chatgpt-export/ directory:

python3 scripts/init_db.py  # Will auto-detect and import

Supports both native ChatGPT format and custom formats.

Phone Contacts (VCF)

Export from your phone (Android: Contacts → Settings → Export):

python3 scripts/import_vcf.py path/to/contacts.vcf

Documents

Automatically indexes all *.md files in memory/ directory.

Advanced Queries

WhatsApp Group Management

-- Find all groups someone is in
SELECT g.name FROM wa_groups g
JOIN wa_memberships m ON m.group_jid = g.jid
WHERE m.phone = '+15551234567';

-- Find contacts in a specific group
SELECT c.phone, c.name FROM contacts c
JOIN wa_memberships m ON m.phone = c.phone
JOIN wa_groups g ON g.jid = m.group_jid
WHERE g.name LIKE '%Family%';

-- Find a group's JID for allowlist
SELECT jid, name, participant_count 
FROM wa_groups WHERE name LIKE '%Project%';

ChatGPT Search

-- Search across all ChatGPT conversations
SELECT c.title, m.content FROM chatgpt_fts f
JOIN chatgpt_messages m ON m.conversation_id = f.conv_id
JOIN chatgpt_conversations c ON c.id = m.conversation_id
WHERE chatgpt_fts MATCH 'project meeting';

Re-indexing

To refresh the database with new data:

rm db/agent.db
python3 scripts/init_db.py

Tips

1. No contact names? WhatsApp only provides phone numbers. Import your phone's VCF to add names.
2. Search not finding? FTS5 uses word boundaries. Use * for prefix matching: Bas* matches "Bashar".
3. Large exports? ChatGPT exports can be 50MB+. First import may take 30-60 seconds.

Implementation Checklist

• Create memory/ directory structure
• Set up daily log template
• Create initial MEMORY.md with sections
• Initialize SQLite database with schema
• Schedule consolidation cron (2-4 AM recommended)
• Add wake-up routine to AGENTS.md
• Test: restart session, verify context loads correctly

Best Practices

1. Write immediately — If it matters, write it now
2. Abstract up — Daily: specific. MEMORY.md: principles
3. Date everything — Context decays; dates help
4. One source of truth — Don't duplicate across files
5. Review regularly — Schedule consolidation, don't skip it

Related

• OpenClaw Documentation
• Cron Scheduling
• Workspace Setup

Inspired by human cognitive architecture. We wake fresh each session, but files are our continuity.

Converted and distributed by TomeVault — claim your Tome and manage your conversions.