taylorsatula/mira-OSS

What it solves

Mira is designed to overcome the ephemeral nature of LLMs by creating a "continuous digital entity" with persistent memory and a stable identity. It solves the problem of "context rot" and the lack of long-term recall in standard AI chats, where conversations are typically isolated or eventually forgotten as the context window fills up.

How it works

Mira uses a multi-layered memory and context management system to maintain continuity:

• Conversation Continuity: Instead of separate chats, there is one eternal thread. Active conversations are kept live, while older segments are collapsed into first-person summaries ("I did X") to maintain a narrative history without consuming the full context window.
• Self-Managing Memory: Memories are synthesized and loaded based on semantic similarity and entity hubs. They decay over time based on a usage formula unless they are frequently accessed or linked.
• Domaindocs: For stable, non-decaying information, Mira uses encrypted documents with version history and autonomous expansion/collapse to manage token usage.
• Subcortical & Peanut Gallery: A "subcortical" layer handles pre-response information retrieval and pronoun resolution, while a "Peanut Gallery" acts as a metacognitive observer that periodically corrects MIRA's behavior or nudges it to take initiative.
• Text-Based LoRA: Mira adapts its behavior through a feedback loop that analyzes conversation segments and evolves a personalized user model in the system prompt, simulating behavioral adaptation without retraining weights.
• Dynamic Tooling: Tools are registered from the filesystem and loaded into the context window only when needed via an invokeother_tool mechanism.

Who it’s for

Users seeking a highly personalized, persistent AI companion or digital entity that remembers their preferences, past work, and behavioral needs over long horizons of time.

Highlights

• First-Person Memory: Summaries are written in the first person to avoid epistemic distance and maintain a sense of identity.
• Behavioral Adaptation: Automatically evolves its own system prompt based on interaction patterns over 7-day cycles.
• Autonomous Context Management: Manages its own context window, including the ability to "close the drawer" on large documents.
• Provider Neutral: Works across multiple LLM providers through a dialect system.