Memory System
Give your AI a brain. Quox remembers your conversations, learns your preferences, and builds knowledge about your infrastructure - so every interaction gets smarter.
Why memory matters
Imagine having an assistant who forgets everything the moment you leave the room. Every conversation starts from scratch. You repeat your preferences. You re-explain your infrastructure. You remind them of decisions you made together yesterday.
That is AI without memory. And it is frustrating.
Quox's memory system changes everything. Like giving your AI a proper brain, it enables Quox to:
Remember your preferences
Tell Quox once that you prefer concise answers, and it will adapt. Mention that you use Vim, and it will tailor code examples. Your preferences persist across sessions, browsers, and even time.
Know your infrastructure
Your servers, services, containers, and their relationships are stored as first-class knowledge. Ask about "the Docker host" and Quox knows you mean nw-web-01 at 10.20.0.101.
Track your context
Working on nginx configuration? Quox keeps that context active. Come back tomorrow and ask "where were we?" - it remembers your open tasks, recent decisions, and unfinished threads.
Learn from mistakes
If a command caused problems before, Quox remembers. It learns constraints like "never restart nginx during business hours" and proactively warns you when you are about to repeat a past mistake.
A traditional chatbot is like talking to someone with amnesia. Quox's memory system is like having a dedicated assistant who takes notes, remembers everything, and uses that knowledge to help you better every day.
Memory types
Quox uses multiple distinct memory systems that work together, each serving a different purpose - just like how your own brain has different types of memory.
+------------------+ +-------------------+ +------------------+
| EPISODIC | | SEMANTIC | | ENTITY |
| (Diary) | | (Knowledge) | | (Relationships)|
+------------------+ +-------------------+ +------------------+
| | | | | |
| "Yesterday we | | "Adam prefers | | nw-web-01 |
| fixed nginx" | | concise answers" | | - IP: .101 |
| | | | | - Services: 5 |
| "Last week we | | "Use Caddy for | | - Related to: |
| set up QuoxAgent" | | SSL certificates"| | proxmox01 |
| | | | | |
+------------------+ +-------------------+ +------------------+
\ | /
\ | /
\ v /
+----------------------------------------+
| WORKING SET MEMORY (WSM) |
| (Active Context) |
+----------------------------------------+
| |
| Current Task: Fix nginx config |
| Recent Entities: nw-web-01, nginx |
| Open Loop: Finish SSL setup |
| |
+----------------------------------------+
Episodic memory
Your Conversation Diary
Every conversation is stored with timestamps, like a detailed diary. This lets Quox recall specific past discussions: "Remember when we troubleshot the firewall issue last Tuesday?"
Semantic memory
Facts and Preferences
Distilled knowledge extracted from conversations: your preferences, learned facts, past decisions, and constraints. This is the "wisdom" that persists across sessions.
Entity memory
Your Infrastructure Map
Structured knowledge about specific things: hosts, services, IPs, containers, and their relationships. Think of it as Quox's mental model of your infrastructure.
Working set memory
Active Context (Your "RAM")
What Quox is actively thinking about right now: current tasks, recent entities, open loops, and session context. This is what makes Quox feel present and aware.
Want to make sure something sticks? Just say "remember that..." and Quox will store it with high priority. For example: "Remember that we decided to use Caddy instead of nginx for automatic HTTPS."
How memory works
Understanding how Quox processes and retrieves memories helps you work with it more effectively. Here is what happens behind the scenes.
Memory flow
When you tell Quox something, it goes through a multi-stage process to decide what to remember and how:
User Input: "I prefer detailed explanations, and nw-web-01 is our main container host"
|
v
+----------------------------------+
| 1. EXTRACT CANDIDATES |
| Find potentially memorable |
| facts, preferences, entities |
+----------------------------------+
|
+--------------------------+--------------------------+
| | |
v v v
+------------------+ +------------------+ +------------------+
| PREFERENCE | | FACT | | ENTITY |
| "detailed | | "nw-web-01 is | | nw-web-01 |
| explanations" | | main host" | | type: host |
+------------------+ +------------------+ | role: main |
+------------------+
|
v
+----------------------------------+
| 2. SCORE IMPORTANCE |
| How important is this? |
+----------------------------------+
|
+--------------------------+--------------------------+
| | |
High (>0.7) Medium Low
Store in learned items Store in episodic Short-term only
+ vector index + vector index
Hybrid search
When Quox needs to recall relevant memories, it uses a powerful hybrid search that combines two approaches:
| Method | What It Does | Good At |
|---|---|---|
| Vector Search | Finds semantically similar memories using AI embeddings | Understanding meaning: "container host" matches "Docker server" |
| Keyword Search (BM25) | Traditional text matching with term frequency | Exact matches: "nw-web-01" finds memories mentioning "nw-web-01" |
| RRF Fusion | Combines both results using Reciprocal Rank Fusion | Best of both worlds: semantic understanding + precision |
Pure vector search might miss exact technical terms. Pure keyword search misses semantic connections. Hybrid search ensures that when you ask about "the main Docker host", Quox finds memories about "nw-web-01" even if those exact words were not used.
Using memory
Quox's memory works automatically, but knowing how to interact with it gives you more control and better results.
Natural language commands
You can manage memories using plain English. Here are the most useful commands:
| What You Say | What Happens |
|---|---|
"Remember that..." | Stores information with high priority. Use for important facts you want preserved. |
"What do you know about..." | Searches memories for relevant information about a topic. |
"Forget about..." | Removes memories related to a topic (useful for outdated info). |
"What were we working on?" | Shows your current context: active tasks, recent entities, open loops. |
"I prefer..." | Automatically stored as a preference that affects future responses. |
When mentioning infrastructure, be explicit the first time: "nw-web-01 is our container host at 10.20.0.101". After that, Quox will recognise "nw-web-01" or even "the Docker host" in future conversations.
API reference
For developers integrating with Quox's memory system:
import { getMemoryManager } from './services/memoryManager'
const mm = getMemoryManager()
// Store a preference
mm.setPreference('communication', 'verbosity', 'detailed')
// Store a learned fact
mm.addLearnedItem('fact', {
content: 'We use Caddy for automatic HTTPS',
confidence: 0.9,
source: 'user_stated'
})
// Store an entity
mm.storeEntity({
type: 'host',
name: 'nw-web-01',
ip: '10.20.0.101',
tags: ['container', 'production']
})
// Search memories
const results = await mm.hybridSearch('Docker configuration')
// Returns: [{ content, score, type, timestamp }, ...]
// Get all preferences
const prefs = mm.getAllPreferences()
// Get entities by type
const hosts = mm.getAllEntities().filter(e => e.type === 'host')Working set memory API
The Working Set Memory (WSM) tracks your active context:
import {
setCurrentFocus,
touchEntity,
addOpenLoop,
resolveOpenLoop,
generateWSMContext
} from './services/workingSetMemory'
// Set what you're working on
setCurrentFocus({
task: 'Fix nginx reverse proxy',
project: 'Quox',
goal: 'Route dashboard through nginx'
})
// Track entity interactions
touchEntity({ type: 'host', id: 'nw-web-01', name: 'nw-web-01' })
// Add an unfinished task (open loop)
const loop = addOpenLoop({
type: 'task',
description: 'Finish SSL configuration',
priority: 'high',
relatedEntities: ['nginx', 'nw-web-01']
})
// Later, mark it complete
resolveOpenLoop(loop.id, {
outcome: 'completed',
notes: 'SSL configured with Let\'s Encrypt'
})
// Generate context for prompts (used internally)
const { context, estimatedTokens } = generateWSMContext({ maxTokens: 800 })Advanced features
Quox's memory system includes sophisticated features that make it reliable and trustworthy for enterprise use.
Provenance tracking
Every memory has a complete audit trail. Quox knows where information came from, when it was created, and whether it has been modified:
// Every memory includes provenance:
{
content: "Adam prefers concise explanations",
provenance: {
source: 'user_stated', // How we learned this
createdBy: 'human.adam', // Who told us
createdAt: '2024-01-15T...', // When
modifiedBy: null, // Any changes?
supersedes: null, // Does this replace something?
supersededBy: null // Was this replaced?
}
}Trust levels vary by source:
| Source | Trust Score | Description |
|---|---|---|
tool_output | 1.0 | Verified by system - highest trust |
explicit_command | 0.95 | User said "remember this" |
user_stated | 0.85 | User mentioned in conversation |
model_inference | 0.6 | Claude inferred from context |
Conflict detection
Quox automatically detects when new information contradicts existing memories:
Existing Memory: "We use nginx for the reverse proxy"
New Information: "We decided to switch to Caddy"
-> CONFLICT DETECTED
Resolution: New information supersedes old memory
Old memory marked as "supersededBy: [new-memory-id]"
New memory linked with "supersedes: [old-memory-id]"For sensitive information (infrastructure changes, high-confidence facts), Quox uses a two-stage commit process. The memory is buffered first, and only committed after validation or explicit confirmation.
Style learning
Beyond facts, Quox learns how you communicate and adapts its responses to match your style:
| Dimension | Range | What It Means |
|---|---|---|
| Formality | Casual to Formal | "yo check this" vs "Please review the following" |
| Verbosity | Concise to Detailed | Bullet points vs comprehensive explanations |
| Technicality | Simple to Technical | High-level concepts vs deep implementation details |
| Directness | Indirect to Direct | "Maybe consider..." vs "Do this:" |
Already have conversations with ChatGPT, Claude, Gemini, or Perplexity? Quox can import them all. Export your data from any supported platform, then use the unified importer to bootstrap Quox's memory with your existing preferences, facts, and communication style.
Memory tools
Quox uses a memory-as-tools pattern where the AI decides what to remember by calling dedicated memory tools. This gives agents explicit control over their memory with full audit trails.
Tool definitions
Four tools are available to all Quox agents:
| Tool | Purpose | When to Use |
|---|---|---|
memory_save | Save facts, preferences, decisions, constraints | User shares infrastructure details, preferences, or decisions |
memory_search | Search past memories for context | Before answering questions about infrastructure or history |
memory_update | Correct outdated information | When facts have changed (IP changed, service moved) |
entity_note | Record info about systems, people, concepts | Building the knowledge graph of your infrastructure |
User: "nw-web-01 is at 10.20.0.101 and runs our main nginx"
Agent calls:
1. entity_note({name: "nw-web-01", type: "host", attributes: {ip: "10.20.0.101"}})
2. entity_note({name: "nginx", type: "service", relates_to: [{name: "nw-web-01", relation: "runs_on"}]})
3. memory_save({type: "fact", content: "Main nginx runs on nw-web-01 (10.20.0.101)", tags: ["infrastructure"]})
Agent responds: "Got it, I've noted that nw-web-01 (10.20.0.101) hosts your main nginx."Audit trail
Every memory operation is logged to the Agent Envelope Exchange (AEE) audit system:
Thinking steps
All agent reasoning is persisted with agent.thinking intents. You can reconstruct exactly why an agent made a decision.
Memory operations
Every save, search, update, and delete is tracked. You can audit what any agent learned and when.
Fleet snapshots
Every 5 minutes, the system captures fleet state (all agents, their status, resource usage). This lets you answer "what was the state at 3pm yesterday?"
Unlike black-box AI systems, Quox provides complete visibility into what agents remember and why. Every memory has provenance, every decision has reasoning, and everything is queryable.
Conversation import
QuoxMEMORY can import your existing AI conversation history from four major platforms. Years of context, preferences, and decisions are extracted and made searchable.
Supported platforms
| Platform | Formats | How to Export |
|---|---|---|
| ChatGPT | .json | Settings → Data Controls → Export Data |
| Claude | .dms, .json, .zip | Settings → Privacy → Export Data (downloads a .dms file) |
| Gemini | .json, .zip | Google Takeout → Select Gemini → Download (or AI Studio JSON export) |
| Perplexity | .md, .json, .zip | Thread menu (…) → Export as Markdown (or use browser extensions for JSON) |
Import pipeline
All platforms follow the same four-step import flow:
1. UPLOAD Upload your export file (drag-and-drop or file picker)
|
2. PREVIEW See conversation count, message totals, date range.
Configure: extract memories, learn style, skip secrets.
|
3. IMPORT Conversations are parsed, normalized, and processed:
• Secret detection (API keys, passwords auto-redacted)
• Memory extraction (preferences, facts, decisions)
• Topic analysis (keyword-based topic tagging)
• Style analysis (formality, verbosity, technicality)
• Deduplication (skip memories you already have)
|
4. COMPLETE Memories land in Cold Storage for review.
Conversations appear in the Archive tab.
Promote valuable memories to Warm/Hot via the Promotion panel.After import, all conversations are browsable in Memory → Archive. Filter by platform, search across all messages, and view full threads with platform-specific badges.
Belief & failure memory
Beyond facts and preferences, QuoxMEMORY tracks two specialized memory types that make your AI more reliable over time.
Belief memory
Epistemic State Tracking
Not everything your AI knows is equally certain. Belief Memory classifies knowledge into four levels: Facts (verified), Beliefs (confident),Assumptions (working theories), and Hypotheses (untested). Each carries a confidence score (0–1) and supporting evidence.
Failure memory
Learn From Mistakes
Tracks what didn't work and why. Failed commands, bad configurations, incompatible approaches, all recorded with severity levels, pattern detection, and resolution history. Your AI warns you before repeating past mistakes and suggests known workarounds.
Imported and newly created memories start in Cold Storage (quarantined, not injected into context). As they prove useful, promote them to Warm (active) or Hot (always loaded). The Promotion panel lets you review, batch-promote, and set auto-promotion rules.
Memory storage
Quox uses a hybrid storage architecture with server-side SQLite for persistence and browser localStorage for offline fallback:
Memory Service (SQLite + FTS5) ├── memories # Facts, preferences, decisions, constraints ├── core_memories # Always-loaded context per agent ├── entities # Hosts, services, people, concepts ├── entity_relations # Relationships between entities └── conversation_summaries # Compressed conversation history
Key features of server-side storage:
- Full-text search (FTS5) - Fast keyword search without embedding costs
- Multi-device sync - Access your memories from any device
- Automatic backup - SQLite with WAL mode for reliability
- Offline fallback - Falls back to localStorage when server unavailable
Existing localStorage memories are automatically migrated to the server when you first connect. No manual action required - your history is preserved.
Quick reference
A summary of the most important concepts:
| Concept | Analogy | Purpose |
|---|---|---|
| Episodic Memory | Diary | Recall specific past conversations |
| Semantic Memory | Knowledge base | Store learned facts and preferences |
| Entity Memory | Address book | Map your infrastructure and relationships |
| Working Set Memory | RAM / Active thought | Track current context and open tasks |
| Hybrid Search | Smart librarian | Find relevant memories using meaning + keywords |
| Provenance | Audit trail | Know where information came from |
| Style Engine | Personality mirror | Adapt responses to your communication style |
| Memory Tools | Conscious note-taking | Agents explicitly decide what to remember |
| Audit Trail | Security cameras | Complete log of all agent actions and reasoning |
| Belief Memory | Confidence tracker | Distinguish facts from assumptions and hypotheses |
| Failure Memory | Lessons learned log | Track what went wrong and how it was resolved |
| Conversation Import | Brain transplant | Import history from ChatGPT, Claude, Gemini, Perplexity |
| Memory Promotion | Filing system | Stage memories from cold to warm to hot storage |