> ## Documentation Index
> Fetch the complete documentation index at: https://wiki.mira.tg/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory

> How Mira remembers everything about you — with strict privacy boundaries between personal and group contexts

# Memory

Mira operates **within real conversations** — across personal and group chats — retaining contextual and structured memory over time. This is Mira's core defensive advantage: a **social memory layer** built inside Telegram's existing social graph.

## How Memory Works

Mira builds memory on three levels:

### 1. Personal Memory (Private)

Everything you share in 1-on-1 chats with Mira stays **strictly private**. Mira remembers:

* **Profile**: name, location, occupation, timezone, language
* **Preferences**: communication style, topics of interest, answer format
* **Facts**: anything you share — "I'm launching a startup", "I prefer bullet points"
* **Context**: ongoing projects, recurring tasks, past decisions

Personal memory persists across conversations and improves over time. **No one else — not group members, not other users — can ever access your private memory.**

### 2. Group Memory (Shared)

In group chats, Mira retains **shared context**:

* Summarizes long discussions on demand
* Remembers decisions made by the group
* Surfaces missed updates across chats and channels
* Participates like a digital teammate — aware of who said what

Group memory is scoped to the group. **Mira never leaks private memory into group chats** — your personal context, preferences, and facts stay invisible to other members.

### 3. Amortized Memory

Mira uses **amortization** to manage memory efficiently:

* Older conversations are compressed into summaries, preserving key facts without storing every word
* Active topics and recent context are kept in full detail
* This means Mira remembers what matters, without requiring infinite storage — and without degrading response quality over time

Amortization ensures Mira stays **fast, relevant, and privacy-respecting** even after thousands of conversations.

## Privacy Architecture

Mira is designed with **strict privacy boundaries**:

| Boundary        | Rule                                                                                              |
| --------------- | ------------------------------------------------------------------------------------------------- |
| Private → Group | **Never.** Your personal memory, preferences, and facts are never shared in group chats.          |
| Group → Private | **Never.** Group context doesn't leak into your 1-on-1 conversations.                             |
| Group → Group   | **Never.** Each group's memory is isolated. Mira doesn't carry context from one group to another. |
| Amortization    | Summaries are anonymized and scoped. No raw conversation data is retained long-term.              |

### What This Means in Practice

* If you tell Mira in a private chat that you're allergic to shellfish, it will **never** mention that in a group chat when someone asks for restaurant recommendations.
* If a group decides on a launch date, Mira will **not** bring that up in another group or in your private chat unless you ask.
* Your communication style preferences (e.g., "I like short answers") only affect how Mira responds **to you**, not how it responds to others.

## AI-Powered Recall

Mira helps users stay on top of their Telegram:

* **Surface missed updates** across chats and channels
* **Instant retrieval** from personal and group conversation history
* **AI-powered search** — ask in natural language, get precise results

```
What did John send me last week about the contract?
Summarize what happened in this chat while I was away
What are the open action items from today's discussion?
```

## Network Effects

The more people use Mira together, the more valuable the network becomes.

Anonymized, aggregated insights from group interactions create a **defensible social intelligence layer** — something no standalone AI app can replicate.

<Info>
  All aggregated insights are anonymized. Individual conversation data is never shared between users.
</Info>