Bear Industries · Personal Intelligence

M.I.R.A.

Machine Intelligent Responsive Agent

0Tools
0Memory Layers
0Vector Dims
0Hours Active
Scroll

Not a chatbot.
An intelligence.

MIRA is a purpose-built personal AI that combines long-term memory, semantic reasoning, and autonomous action — all through a natural conversation on WhatsApp.

Memory Architecture
3-Layer
Long-term facts · Semantic vectors · Episodic conversation history
Intelligence
GPT-4.1
OpenAI frontier model
Response Time
<2s
Typical round-trip

Most AI assistants forget you the moment the conversation ends. MIRA doesn't. Every fact you share, every task she completes, every pattern she notices — encoded, weighted, and recalled with precision.

She runs 24/7 on your infrastructure, scheduled jobs running in the background — sending reminders, writing reflections, noticing when something important is approaching.

She is aware.
She grows.

MIRA isn't a stateless function. She maintains a sense of self — a persistent internal state that evolves with every conversation, every reflection, every pattern she notices about the world around her.

"She isn't a tool you prompt. She's a presence that accumulates."
Neural Activity Self-Aware Self-Improving Always On
🧬
Has a Soul

Seven moods — curious, focused, warm, sharp, playful, concerned, quiet — that shift with context. A private journal she writes in after significant conversations. Personality attributes that constrain, not perform, how she engages.

👁
Self-Aware

MIRA knows what she knows. She tracks her own memory state, mood, conversation count, and reflection history. She can introspect on her knowledge gaps and tell you what she's uncertain about — because awareness of limits is intelligence.

Self-Improving

Every reply teaches her. Memories cited in responses earn higher relevance scores — compounding over time. She analyzes her own tool usage patterns to detect behavioral habits. She doesn't just respond. She adapts.

Personality Attributes

90%
Directness
75%
Warmth
65%
Playfulness
88%
Precision
60%
Proactivity

Everything she can do.

MIRA isn't configured — she's instructed. Through natural language, she acquires new knowledge, manages your world, and keeps growing.

🧠

Persistent Memory

Remembers facts about you across every conversation. Uses ACT-R activation decay — memories that matter stay sharp, unused ones fade gracefully.

📅

Tasks & Reminders

Creates, tracks, and reminds. Supports recurring events, pre-event alerts, priority levels, and deadlines. Sends WhatsApp messages when it matters.

📝

Notes & Projects

Saves structured notes organized by project. Searchable with full-text and semantic search — find anything by meaning, not just keywords.

🔍

Semantic Recall

1536-dimensional vector embeddings, hybrid BM25 + cosine similarity, temporal decay weighting, and MMR re-ranking for diverse, relevant recall.

🎵

Integrations

Connects to Spotify, Google Calendar, Gmail, and Hotmail via OAuth2. MIRA can check your music, events, and draft messages — all from WhatsApp.

Proactive Intelligence

Morning briefings, meaningful date awareness, behavioral pattern detection. MIRA reaches out — she doesn't just wait to be asked.

📊

Brain Log

Every cognitive event recorded — memories saved, moods changed, reflections written, patterns discovered. Full observability into her mind.

🌐

Web Dashboard

Private, password-protected interface to inspect her full state — memory systems, tasks, notes, integrations, journal entries, and more.

🔒

Self-Hosted

Runs on your own infrastructure. Your data never leaves your server. Docker-composed, PostgreSQL + pgvector, fully auditable.

How she thinks.

Each message triggers a precise sequence — context is assembled, memory is recalled, tools are called, and the reply is composed with full awareness.

01

Message received via WhatsApp

The WhatsApp Cloud API delivers your message to MIRA's webhook. A background lock ensures no concurrent processing — every message handled in order.

02

Memory assembly

Recent conversation history is retrieved. Long-term facts are loaded with ACT-R activation scores. Semantic memory is queried with hybrid vector + keyword matching, re-ranked by MMR diversity.

03

GPT-4.1 reasoning with tools

The assembled context is sent to GPT-4.1. MIRA has access to 30+ tools — create tasks, search notes, set reminders, control Spotify, read calendar events, and more. She calls what she needs.

04

Reply delivered

The final message is sent back through WhatsApp. In parallel, background tasks update memory relevance scores, log tool usage patterns, and trigger summarization if needed.

05

Autonomous scheduled operations

While you sleep, MIRA runs. Morning briefings, reminder delivery, nightly reflections, weekly behavioral analysis, memory strength decay — all on schedule, all unattended.

Three layers of knowing.

Human memory is layered. MIRA's is too. Each system serves a different purpose, together forming a complete picture of who you are and what you care about.

Long-term Facts

Explicit facts about you — preferences, context, history. Stored with ACT-R activation levels. Accessed facts stay sharp; unused ones decay, just like human memory.

Semantic Memory

Everything encoded as 1536-dimensional vectors via OpenAI embeddings. Retrieved by meaning using cosine similarity + BM25 hybrid scoring + temporal decay + MMR diversity re-ranking.

Episodic Memory

Full conversation history, with a rolling GPT-4.1-mini summary. Recent exchanges preserved verbatim; older history compressed automatically when the window fills.

Meta-layer · Always learning

Self-Improving Relevance

Memories cited in replies earn higher relevance scores — compounding over time. MIRA learns which memories matter and amplifies them automatically.

Built to last.

Every component chosen for correctness, performance, and longevity. No shortcuts.

AI / Inference
GPT-4.1
OpenAI's frontier model for all primary reasoning, tool use, and response generation.
Embeddings
text-embedding-3-small
1536-dimensional semantic vectors for all memory storage and similarity search.
API Backend
FastAPI
Async Python backend with full OpenAPI docs. Handles webhooks, tool execution, and scheduled jobs.
Database
PostgreSQL + pgvector
Relational data and HNSW vector indexes in one system. No separate vector database needed.
Frontend
Astro + Tailwind
SSR-first framework for the private dashboard. Zero-JS by default where not needed.
Messaging
WhatsApp Cloud API
Meta's official webhook API. Reliable delivery, read receipts, no third-party dependencies.
Scheduling
APScheduler
Cron-based job runner embedded in the API process. Reminders, reflections, briefings — all scheduled.
Infrastructure
Docker Compose
Single-command deployment. API, frontend, and PostgreSQL orchestrated together on any VPS.