Skip to content

Memory & Context

Queria manages memory at multiple levels to offer you a personalized conversation experience. The system learns who you are, what you do and how you work - and uses this information to improve every answer.

4-level architecture

The memory system operates across four integrated levels:

LevelWhat it doesWhere it lives
ExtractionCollects information about youChat, documents, behavior
ConsolidationImproves memory qualityNightly analysis, weekly review
RetrievalSelects relevant memories for each querySmart relevance scoring
ObservabilityMeasures and improves the systemFeedback, metrics, auto cleanup

Session memory

Within a single conversation, the AI maintains full context. The system automatically tracks:

  • Key facts: data, figures, dates mentioned in the conversation
  • Entities: people, organizations, documents referenced
  • Time scope: the period under discussion
  • Open questions: aspects still to be deepened

This means you can naturally refer to previous messages, ask follow-up questions and build progressive reasoning without repeating context.

TIP

Important facts from the session are automatically promoted to persistent memory after 10+ turns, so they're not lost when you close the chat.

Persistent memory (cross-session)

Beyond the current session, Queria stores information about you that persists between conversations. The system learns from 4 different sources:

1. From your conversations

After each chat turn, an automatic analysis extracts persistent information:

  • Skills: your expertise level in various domains
  • Recurring topics: themes you return to often
  • Key facts: information about your professional context
  • Active projects: what you're working on
  • Preferences: how you prefer to receive answers
  • Interaction style: how you communicate

2. From your documents

When you upload documents, the system analyzes the content to understand what you work on. For example, if you upload insurance contracts, the system understands you work in insurance.

INFO

Document profiling only works for Editor and Reader users, not for Admins managing documents on behalf of others.

3. From your behavior

The system automatically notices your usage patterns:

  • Frequent searches: if you often search "unfair clause", the system memorizes it
  • Documents consulted: documents you open multiple times indicate your interests
  • Filters used: if you always use the "LABOR" filter, the system records it

4. From session promotion

Important facts emerging during conversations are automatically evaluated for promotion to permanent memory. Only truly persistent facts (true tomorrow too) are saved.

How the system improves over time

Nightly analysis

Every night the system analyzes your recent conversations to find patterns that emerge only by looking at multiple conversations together. E.g.: if across 5 different conversations you asked about IVASS regulation, the system understands it's a recurring interest.

Weekly review

Every week, the system re-evaluates older memories:

  • Still-relevant memories: confirmed (confidence rises)
  • Likely outdated memories: demoted (confidence falls)
  • Clearly obsolete memories: removed

This prevents old information from polluting your answers.

Feedback that improves memory

When you give a thumbs up to an answer, the memories that contributed to that answer are reinforced. When you give a thumbs down, they are penalized. This way the system learns which memories are useful and which are not.

Memory and answers

Smart scoring

Not all memories are used in every answer. The system selects the most relevant with a score based on three factors:

FactorWeightWhat it measures
Relevance to question40%How pertinent the memory is to the current query
Confidence30%How sure the system is that the memory is correct
Recency30%How recently the memory was used/confirmed

Contextual suggestions

When your competencies are pertinent to the question, the system suggests connections naturally. E.g., if you work with insurance contracts and ask about a liability clause, the AI might suggest comparing it with the IVASS Regulation you often consult.

"Do you know who I am?"

You can ask the AI what it knows about you at any time. Try questions like:

  • "Do you know who I am?"
  • "What do you know about me?"
  • "What do I work on?"
  • "What are my recurring topics?"

The AI will reply based exclusively on the memories it has collected from your interactions.

Personalized follow-ups

After each answer, follow-up suggestions take three levels into account:

  1. Current answer: aspects to deepen in what was just discussed
  2. Conversation: facts emerged in the previous turns of the same session
  3. Your profile: at most 1 personalized suggestion based on your interests, only if pertinent to the current topic

Memory management

AI Memory page

From the AI Memory page (side menu) you can:

  • View all saved memories by category
  • Delete individual memories that are no longer accurate
  • Delete all memories to start over
  • Disable memory completely (on/off toggle)

Disabling memory

If you prefer the system to remember nothing about you:

  1. Go to AI Memory from the side menu
  2. Disable the "Memory active" toggle
  3. The system will no longer extract memories from conversations
  4. Existing memories remain until you delete them manually

Privacy and isolation

Queria's memory is designed with privacy as a priority:

  • Per-user isolation: each user has their own memory, completely separated
  • Per-company isolation: data from one company is never accessible from another
  • No sharing: your conversations and preferences are not shared with other users
  • Full control: you can delete your memories at any time
  • Transparency: you can see exactly what the system remembers about you

TIP

To get the most from memory, use the platform normally. The system learns automatically from your usage patterns. The more you interact, the more answers will be personalized for you.


Queria v3.5.0 -- Cog-RAG Architecture

Queria - Document Intelligence con Cog-RAG