taylorsatula/mira-OSS

This is the public release of MIRA OS. Discrete memories decay through momentum loss, tools auto-configure when dropped into tools/ folder, and the system prompt composes from modular trinkets. I would like to think I've made an elegant brain-in-box. You load it and send cURL requests - it talks back, learns, and uses tools. Contributions welcome.

What it solves

MIRA is designed to create a continuous digital entity with persistent memory and recall, overcoming the ephemeral nature of standard LLM context windows. It eliminates the "start a new chat" functionality, forcing the system to manage its own history and identity over the long term.

How it works

MIRA uses a multi-layered memory architecture to maintain continuity:

  • First-Person Memory: Instead of third-person logs, MIRA generates first-person summaries of conversation segments that collapse into "memories" with absolute timestamps to maintain a sense of self and narrative continuity.
  • Self-Managed Memory: Memories decay over time based on a usage formula unless they are frequently accessed or linked, preventing context rot without requiring manual curation.
  • Domaindocs: Stable, encrypted documents for long-form text that MIRA can autonomously expand or collapse to save tokens.
  • Subcortical Layer: A pre-response information-retrieval pass that resolves pronouns, extracts entities, and prepares the necessary context before the primary model generates an answer.
  • Peanut Gallery: A metacognitive observer that monitors the conversation every five turns to provide coaching or corrections if the interaction goes off track.
  • Text-Based LoRA: A behavioral adaptation loop that analyzes conversation patterns and folds them into a descriptive user model in the system prompt every seven use-days.

Who it’s for

Users who want a persistent, evolving AI companion or agent that remembers their preferences, work history, and specific behavioral needs without the need to manually manage chat histories or separate knowledge bases.

Highlights

  • Continuous Conversation: A single, forever-thread approach to interaction.
  • Agentic Tool Use: Tools are discovered from the filesystem and can beled dynamically by MIRA using an invokeother_tool to avoid token waste.
  • Provider Neutrality: Works across multiple LLM providers via a dialect system.
  • Automated Context Management: Combines segment collapse, live compaction, and semantic similarity for efficient memory retrieval.