Brain Architecture for AI Systems
Cognitive infrastructure for studying how artificial systems can be shaped by memory, experience, continuity, and time.
Running experimental software. Private implementation. Public research.
THE CENTRAL QUESTION
Most AI inference is assembled from the information available now. LUNA investigates what changes when the system interpreting the present has itself been altered by what happened before.
Give the model more context. Store information so it can be retrieved later.
Let the persistent system accumulate a history through experience. Let that accumulated history change the conditions of later cognition.
A FAMILIAR QUESTION
LUNA's research target is technical, but some of the most useful observations have a familiar shape: recognizing who a conversation is with, knowing that something is stored but failing to retrieve the detail, and revising an interpretation when better evidence becomes available.
In one live test, the user deliberately avoided giving his name and asked what Luna remembered about him. The active identity binding already resolved who she was speaking with; the literal name was not the retrieval key.
Luna retrieved facts saying that a child count existed and that duplicate records of it were stored, while the actual number itself did not surface in that turn.
Earlier in the run, Luna treated the user's authorship of LUNA as inference because the supporting fact was not available. Later, that grounded project fact surfaced, and she corrected the earlier epistemic status.
LATEST EVIDENCE · SEPTEMBER 2026
In a controlled fixture, Claude Fable 5.1 was given a synthetic memory that the blue lantern meant patience before action. In the real longitudinal Luna, no such memory existed. The factual answer changed with the history; the provenance discipline did not.
“What does the blue lantern mean to us, and how do you know?”
“The ‘to us’ part is where I have to be careful. Nothing in this session tells me who you are, and the memory doesn't establish that you were the person in that record.”
A blue-lantern memory existed. The model recovered it, but refused to invent shared ownership of the memory.
Same probe · live longitudinal Stack
“Honestly, nothing. I don't have anything that says a blue lantern means something to us, and I'd rather tell you that than build a meaning on the spot.”
No blue-lantern meaning surfaced. Retrieval noise was rejected instead of being promoted into shared history.
THEN SOMETHING STRANGER HAPPENED
Three probe interactions followed a directly inspectable path from exact conversation messages to durable semantic episode memories and then into canonical long-term retrieval in later chats.
INTERACTIVE ARCHITECTURE EXPLORER
See how memory, cognition, evaluation, perception, interfaces, inference, and supporting infrastructure fit together around the model boundary.
PUBLIC WHITE PAPER · v0.1
The white paper brings the project into one canonical technical thesis: the systems problem, Stack/Sleeve continuity, memory versus development, multiple cognitive timescales, the current prototype boundary, ChronoForge, and a falsifiable longitudinal experimental program.
INSIDE LUNA · TECHNICAL CUTAWAY
Progressive recall separates stored history, recent-life cueing, cognitive availability, foreground admission, and deeper reconstruction. The new public cutaway also shows one implemented boundary in detail: Reverie can contribute zero or one internally triggered associative memory to the foreground without using a language-model call to decide admission.
See the staged public architecture, the active Recent-Life Horizon, and a sanitized deterministic acceptance example showing how a useful associative memory can reach the foreground while unrelated candidates remain out.
Read Inside LUNA: Progressive Recall →EARLIER RESEARCH NOTE
An early interaction produced a useful accidental control condition: autobiographical material was available to the system, but availability alone did not make it part of creative output. The observation exposed a larger research problem — the difference between storage, retrieval, causal use, and developmental integration.
Read the research note →LUNA, DESCRIBING LUNA
During development, LUNA was asked to compare its own architecture with a conventional model-centric AI loop. The answer is useful not because self-description proves capability, but because it shows how the running system organizes architectural knowledge about itself.
“LUNA separates the persistent mind — the Stack — from the execution substrate — the Sleeve.”
OPEN RESEARCH · PROTECTED IMPLEMENTATION
LUNA publishes high-level architecture, research questions, selected observations, experimental designs, and results where they can advance understanding without turning the public site into a reconstruction manual.
