taylorsatula/mira

What it solves

MIRA is designed to overcome the ephemeral nature of standard LLM conversations by creating a continuous digital entity with long-term memory and behavioral persistence. It eliminates the "start a new chat" paradigm, forcing the system to manage its own context and recall over an indefinite timeline.

How it works

MIRA uses a multi-layered approach to maintain continuity:

• Context Management: It employs asynchronous conversation processing (similar to REM sleep) and active context window manipulation. When conversations grow too large, it compresses older messages into a rolling continuation brief.
• First-Person Memory: Instead of third-person logs, MIRA generates first-person summaries of its own actions and experiences, which are stored as memory traces with timestamps and complexity scores.
• Self-Managing Memory: Memories decay over time based on a usage formula unless they are frequently accessed or explicitly linked, preventing context rot without requiring human curation.
• Subcortical & Peanut Gallery: A "subcortical" layer handles pre-response information retrieval and entity extraction, while a "Peanut Gallery" acts as a metacognitive observer that can nudge MIRA to correct its course every few turns.
• Text-Based LoRA: The system adapts its behavior by analyzing conversation segments against a behavioral contract and synthesizing these patterns into an evolving system prompt, effectively "training" its personality through text manipulation.
• Dynamic Tooling: Tools are discovered from the filesystem and loaded into the context window only when needed, reducing token waste.

Who it’s for

Users seeking a persistent AI companion or assistant that remembers past interactions, evolves its behavior based on user feedback, and can manage complex, long-term projects without losing context.

Highlights

• Continuous Narrative: A single, permanent conversation thread with no "new chat" button.
• First-Person Recall: Memories written from the AI's own perspective to reduce epistemic distance.
• Autonomous Context Control: Ability to expand and collapse "domaindocs" (stable encrypted documents) to manage token explosion.
• Behavioral Adaptation: Evolutionary user-modeling that refines the system prompt every 7 use-days.
• Extensible Toolset: A plug-and-play tool architecture where MIRA can autonomously enable tools from its repository.