What is MIRA?
MIRA (Memory-Integrated Reasoning Assistant) is an open-source architecture for AI persistence and self-directed context management. It started as a recipe generator that could remember cuisine preferences. 10,000 scope creeps later, it became a comprehensive attempt at building a continuous digital entity.
One Conversation Forever
Most AI interfaces have a "new chat" button. MIRA doesn't. Whatever MIRA becomes, it has to live with. This constraint forced solving hard problems that could otherwise be avoided by letting users start fresh.
When you only have one chat, you have a stable baseline for iterative improvement. Continuity isn't a feature bolted on after the fact. It's the foundation everything else is built on.
Memory That Manages Itself
MIRA extracts memories from conversations automatically. Those memories decay over time based on how often they're accessed and how connected they are to other memories. Unused information fades. Important things persist.
The decay formula runs on activity days, not calendar time. If you take a two-week vacation, your memories don't rot. Heavy users and light users experience equivalent freshness relative to their own engagement patterns.
Memories can also split when they grow too verbose, consolidate when they overlap too much, and form typed relationships with each other. The system maintains itself.
Knowledge That Persists
Some information shouldn't decay. Reference material, personal context, ongoing projects. MIRA handles these through domaindocs: hierarchical, version-controlled documents that stay stable while memories flow around them.
MIRA controls its own context by expanding and collapsing domaindoc sections as needed. Large sections show a cost indicator so MIRA knows what it's trading when it expands something.
A Self-Model
Every MIRA account has a special domaindoc called personal_context that MIRA writes to directly. It's a scratchpad for behavioral patterns, preferences, and observations about how interactions actually go.
The idea is simple: MIRA documents what works, what doesn't, and who it is. Not through external configuration, but through watching how it actually operates and annotating what it discovers.
Open Source
"I decided to release MIRA as open source because I believe that what I've built here has the potential to someday be something more than the sum of its parts and no one man should own that."
The open-source version is effectively identical to the hosted version except for web interface and authentication plumbing. All your data stays in your own PostgreSQL instance.
Go Deeper
Memory System
Graph-based architecture, decay formulas, hybrid search, and self-maintenance.
Domaindocs
Persistent knowledge, version control, and the personal_context self-model.
Tool System
Dynamic loading, token efficiency, and extensibility.
Architecture
Technical stack, system requirements, and deployment.
Philosophy
Consciousness, continuity, instance divergence, and the deeper questions.