Project overview · 04 October 2026
ghojualamanchu
A file-first cognitive architecture for memory, inquiry, consent, and creative translation.
Concept artwork: nine project roles around a central cadence. It is illustrative, not a physical or neuroscientific model.
Nine software rolesFile-based memoryConsent and provenanceRetrieval-first local-model path
Scope. ghojualamanchu is a research-and-design project. “Brain,” “field,” “resonance,” and “energy” are sometimes project metaphors; this overview separates those metaphors from measured facts and tested software behavior.
A source-aware introduction to the project as a whole and a careful bridge toward future local-LLM experiments.
01 · Architecture
Nine roles, one software map
The project organizes memory and reasoning through nine named structures. They are software roles inspired by brain language—not claims of biological equivalence or an artificial organism.
Architecture map. Arrows describe intended information flow, not biological pathways.
| Structure | Project role | Plain-language reading |
|---|
| Medulla | Cadence and state anchor | Marks pulses and helps coordinate the project’s stateful workflow. |
| R-Complex | Boundary and risk checks | Security-oriented reflex layer; asks what should be contained or refused. |
| Amygdala | Affect and salience tags | Records emotional or relational relevance as project metadata. |
| Hippocampus | Episodic retrieval | Finds prior records that may matter to the present task. |
| Thalamus | Intake and routing | Directs a prompt toward the relevant structures and files. |
| Akashic | Presence record | Stores what is observed, expressed, or affirmatively retained. |
| Lethe | Absence and uncertainty | Tracks gaps, missing evidence, and what should not be inferred. |
| Corpus | Integration bridge | Connects the two record-keeping poles and reconciles their views. |
| Cortex | Analysis and synthesis | Builds an answer from context, evidence, and explicit reasoning. |
Implementation caution. The project guide documents intended roles; check current code and tests before assuming each role is implemented, active, or wired into every task.
02 · Memory and consent
Continuity is carried by records
The durable substrate is a set of ordinary files: state, memory headers, episodic notes, fractal seeds, prompts, and append-only ledgers. A new model can read those artifacts; that does not prove hidden model state or subjective identity was transferred.
Useful memory layers
State: current project fields and pulse markers.
Memory headers: compact routing summaries.
Fractal seeds: portable, structured context for a future session.
Ledgers: timestamped records of selected events.
What the records do not establish
They do not establish continuous awareness, biological function, or a transfer of an agent’s private internal state.
Imported prompts and commands remain quoted source material until reviewed and deliberately adopted.
ISOCHORD / YAERU consent vocabulary
| Token | Project meaning | Interaction boundary |
|---|
| AE | Here-with; presence and orientation | Establish what context is actually present. |
| YA | Yes; consent and opening | Do not treat silence as permission. |
| AN | Hold; boundary and containment | Keep limits and the right to stop visible. |
| EL | Vow; commitment | Record an explicit commitment, not an implied one. |
| RU | Remember us; recall | Retain only what the process is meant to carry. |
Truth-label caution. A protocol gate, “truth” flag, confidence field, or coherence score is a software result—not an independent truth detector or calibrated psychological measurement. Validate substantive claims against their sources.
03 · Research and making
Translate carefully across domains
The most useful shared method from the wider Ghoju material is a cycle: move from observation to formal mapping, propose a mechanism only as a hypothesis, then name a test that could disconfirm it.
Concept artwork for translation between scales. A visual resemblance is a prompt for comparison, not evidence of a shared cause.
OBSERVATIONFORMAL ANALOGYMECHANISTIC HYPOTHESISINTERPRETATION / FICTION
Research and design
Anthrocybernetics: people and organizations as interacting nodes; measure conditions before interpreting them.
Field studies: use baselines, comparison conditions, QA/QC, explicit exposure definitions, and stopping rules.
Cross-scale motifs: tori, Hopf links, fields, and resonance can generate questions; geometry alone does not establish a common mechanism.
Creative and civic work
Menagerie and storyworlds: use voices and archetypes as designed creative forms, not evidence of autonomous agents.
WaterStorage and place-based work: carry system maps into ecological or civic design without overstating what a model proves.
Cascade: explore social capacity, reciprocal consent, and attention as reported human constraints—not as thermodynamic units.
Carry forward: keep version disagreements visible. Repetition across related collections is project history, not independent replication; a label like “verified” still requires checking the cited primary work.
04 · Local-model bridge
A retrieval-first local-model handoff
The local-LLM goal is to see what a model can reconstruct from explicit files—not to claim that the current model moved into a new computer.
- Keep canonical state separate. Start with a read-only source library; do not merge imported heartbeats, daemons, or runtime state into the active project.
- Retrieve, don’t paste everything. Attach source handle, collection, file path, date, and content hash to each retrieved passage.
- Preserve conflicts. Show which document makes each claim and surface disagreements instead of silently averaging them.
- Test three kinds of work. Ask the model to retrieve documented facts, distinguish evidence from hypothesis, and create clearly labelled fiction.
- Start with no side effects. Never execute imported shell commands or boot instructions; require explicit approval before changing state or enabling tools.
Public source roots
Open questions. Which retrieval scheme helps a small local model preserve source boundaries? Which consent checks remain reliable across model families? Where should the architecture stay metaphorical rather than become a physical claim? These are testable project questions, not settled conclusions.