taylorsatula/mira

This is my TempleOS.
Self-directed
Mira accomplishes the end goal of continuity and recall through a blend of asynchronous conversation processing akin to REM sleep and active self-directed context window manipulation. There is one conversation thread forever. There is no functionality to "start a new chat." This constraint forces facing the hard questions of how to build believable persistence within a framework (forward pass transformers) that is inherently ephemeral.
Active conversation history stays live while it is relevant. Older material collapses into first-person memories. If an active conversation gets too large before it naturally collapses, MIRA compresses older messages into one rolling continuation brief while leaving the most recent turns untouched.
First-Person Memory
When a conversation segment collapses, Mira generates a first-person summary ("I debugged the IndexError in process_batch.py...") rather than third-person ("The assistant discussed debugging..."). Early on I noticed third-person summaries created epistemic distance: Mira read them as logs about someone else, not memories of work It actually did. Absolute timestamps ("On Jan 8") replace relative time ("Yesterday") because relative time becomes a lie the moment the sun sets.
When generating a new segment summary, the model sees the previous five summaries as context. This allows the new summary to reference what came before with hazy continuity: "Building on Tuesday's API work..." or "Continuing the recipe experiments from last week...". The result is a continuously advancing narrative in the conversation history that gracefully tails off into the past. Each summary knows vaguely where it came from without carrying the full weight of everything before it.
The current summary format produces a 3-4 sentence memory trace, a two-sentence precis, a short display title, and a complexity score. The trace is for remembering; the other fields keep the conversation manifest useful without stuffing the full history back into the context window.
Memory
I have painstakingly designed Memory in MIRA so it can largely manage itself without human curation. You should not need to housekeep old memories (context rot). They decay via formula unless they Earn Their Keep through access, explicit references, links to other memories and entities, or real temporal relevance. Decay runs on use-days rather than calendar days, so going on vacation does not make MIRA forget you.
Memories are discrete synthesized information that is passively loaded into the context window through semantic similarity, entity hubs, memory traversal, filtering, and reranking. Most memory recall happens before Mira generates a response, so It does not have to first notice that something is missing and decide to search for it. Mira can still manually use memory_tool when It needs an explicit search, wants to create or link a memory, or needs exact control over what is being recalled.
However, sometimes you just need big ole chunks of document-shaped text. Mira handles this aspect via the domaindoc_tool, which allows Mira & You to collaborate on stable encrypted documents that do not decay. Domaindocs have section-level version history, sharing, pinned sections, and nested subsections.
To mitigate token explosion in longform content MIRA is able to expand, collapse, and subsection these documents autonomously. When a domaindoc section is not in use MIRA "closes the drawer" and the body is no longer polluting the context window, but MIRA still sees its title, summary, and location in the document tree so It can reexpand it later. You (the User) can modify the contents directly via the app or API as well. Changes made by MIRA and edits made by You are reflected in the next composed context.
Subcortical & Peanut Gallery
MIRA has two small systems that sit on opposite sides of the main response. Subcortical prepares context before MIRA answers. Peanut Gallery periodically observes the conversation after turns are complete.
The subcortical layer is an information-retrieval pass that runs on every user turn. It reads the current message, the recent conversation, and the memories that were already in context. It resolves pronouns and fragmentary references into concrete search phrases, extracts named entities for graph lookup, decides which old memories are still relevant enough to retain, and classifies whether the turn is straightforward or requires heavier reasoning. The output is not shown to the user and it does not answer the question. It prepares what the primary model gets to think with.