MIRA
A self-hosted, single-conversation AI agent framework that keeps one continuing conversation thread instead of separate chats, using asynchronous processing and self-directed context management to build long-term continuity and memory.
MIRA has no functionality to start a new chat; active conversation history stays live while relevant, and older material collapses into first-person memory summaries rather than being dropped, written as "I debugged the IndexError..." instead of a third-person log, with absolute timestamps rather than relative ones such as "yesterday." If an active conversation grows too large before it naturally collapses, older messages compress into one rolling continuation brief while the most recent turns stay untouched.
When a new segment summary is generated, the model sees the previous five summaries as context, so the new one can reference what came before without carrying the full prior history; each summary produces a memory trace, a two-sentence precis, a short display title and a complexity score. Memories decay by a formula unless renewed through access, so they do not need manual housekeeping.
It is written in Python, uses PostgreSQL with pgvector for storage, integrates with Anthropic's Claude, and is released under the AGPL-3.0 license.