Advanced Memory

Vector-powered memory with semantic search, style learning, and unlimited storage.

Overview

Advanced Memory is a premium plugin that upgrades QuoxCORE's memory system from basic keyword storage to a full vector-powered memory engine. It removes the free tier's storage limits, enables semantic search through Qdrant, and adds capabilities such as style learning, belief tracking, and conversation import.

The free memory tier stores up to 10 facts and 3 preferences with keyword-only search. Advanced Memory removes all limits and switches to meaning-based retrieval.

Tier: Pro Price: $99 one-time purchase

Vector Memory (Qdrant)

Advanced Memory integrates with Qdrant, a vector database, to store and retrieve memories by semantic similarity rather than keyword matching.

How It Works

When a memory is stored, its text content is converted to a 1,536-dimension embedding vector using OpenAI's embedding model.
The vector is stored in a Qdrant collection alongside the original text and metadata.
When searching, the query is also converted to a vector and compared against stored memories using cosine similarity.
Results are ranked by semantic relevance, not keyword overlap.

Configuration

Setting	Default	Range	Description
`semanticSearchThreshold`	0.6	0.3 - 0.9	Minimum similarity score for results. Lower values return more results with less precision.
`contextTokenLimit`	4,000	500 - 8,000	Maximum tokens injected from memory into each conversation.
`retentionDays`	90	7 - 365	Days before episodic memories are archived.

The Qdrant instance runs alongside QuoxCORE on port 6333. No external service is required.

Semantic Search

Semantic search finds memories by meaning rather than exact keyword matches. Examples:

Query	Finds memories about...
"deployment process"	CI/CD pipelines, release procedures, Docker builds
"the database issue last week"	PostgreSQL errors, migration failures, connection timeouts
"how does authentication work"	OAuth flows, API keys, session management

The search engine combines vector similarity with metadata filtering. You can search across all memory types or filter by specific categories (facts, preferences, decisions, entities).

RAG (Retrieval-Augmented Generation)

Every conversation automatically retrieves relevant memories and injects them into the context window. This happens transparently:

The user's message is embedded as a vector.
The top-matching memories are retrieved from Qdrant.
Matching memories are formatted and prepended to the conversation context.
CommanderQ responds with awareness of the retrieved context.

The free tier injects up to 500 tokens of memory context. Advanced Memory increases this to 4,000+ tokens, providing substantially more background knowledge in every response.

Memory Types

Advanced Memory supports all QuoxCORE memory types with no storage limits:

Facts

Declarative knowledge items. Examples: "nginx runs on nw-web-01", "the staging environment uses PostgreSQL 15", "deployments happen on Tuesdays".

Preferences

User communication and workflow preferences. Examples: preferred verbosity level, code style conventions, notification settings.

Decisions

Records of choices made and their rationale. Useful for maintaining consistency and explaining past reasoning.

Entities

Structured records for hosts, services, people, and projects. Each entity has a type, name, and associated properties. Entity memory powers features like the fleet directory and project context.

Beliefs and Rationale

Tracks the distinction between confirmed facts, working assumptions, and hypotheses. Each belief records its confidence level and the evidence supporting it. This allows CommanderQ to communicate uncertainty accurately.

Failure Memory

Records what did not work and why. When a similar situation arises, CommanderQ can warn you before repeating a known mistake. Each failure record includes the attempted action, the outcome, and the lesson learned.

Working Set Memory (WSM)

Tracks active context across sessions:

Current focus — what you are working on right now
Recent entities — hosts, services, and projects you have interacted with recently
Open loops — unresolved tasks and threads that need follow-up

WSM ensures that when you return to a conversation after hours or days, CommanderQ remembers what was in progress.

Style Engine

The Style Engine learns how you communicate and adapts CommanderQ's responses to match. It tracks:

Formality — whether you prefer formal or casual language
Verbosity — whether you want concise answers or detailed explanations
Technical depth — whether you prefer high-level summaries or low-level detail
Terminology — domain-specific terms you use frequently

Style learning is automatic. The engine analyses your messages over time and gradually adjusts its model of your communication preferences. You can enable or disable it via the enableStyleEngine configuration option.

ChatGPT Import

Import your entire ChatGPT conversation history into QuoxCORE. The importer:

Reads your exported ChatGPT data (JSON format from OpenAI's data export)
Extracts facts, preferences, decisions, and entities from conversation content
Generates vector embeddings for all extracted memories
Stores everything in the Qdrant collection, searchable immediately

This transfers years of accumulated context into QuoxCORE in minutes. All imported memories are tagged with their source for filtering.

Contradiction Detection

When new memories conflict with existing ones, the contradiction detector flags the conflict and asks for resolution. For example:

Existing memory: "deployments happen on Tuesdays"
New memory: "we moved deployments to Thursdays"

The detector identifies the contradiction and either updates the existing memory or creates a superseding record, maintaining a consistent knowledge base.

Memory Visualisation

Premium users can view their memory store through three visualisation modes:

Timeline — memories plotted chronologically, showing when knowledge was acquired
Knowledge graph — entity relationships displayed as a network diagram
Heatmap — memory access frequency, showing which knowledge is retrieved most often

Free vs Premium

Capability	Free	Premium
Memory storage	10 facts	Unlimited
Preferences	3	Unlimited
Context window	500 tokens	4,000+ tokens
Search	Keyword only	Semantic (vector)
Style Engine	Not available	Automatic
Belief memory	Not available	Full support
Failure memory	Not available	Full support
ChatGPT import	Not available	Full support
Contradiction detection	Not available	Automatic
Memory visualisation	Not available	Timeline, graph, heatmap

Requirements

QuoxCORE 1.0.0 or later
QuoxAgent 1.0.0 or later
Qdrant vector database (included with QuoxCORE Docker deployment)

Activation

Activate through the QuoxCORE plugin settings page. The licence is validated against the QuoxAgent instance. Once activated, all storage limits are removed and semantic search is enabled.