| --- |
| license: other |
| license_name: ameforge-proprietary |
| license_link: LICENSE |
| language: |
| - en |
| pipeline_tag: question-answering |
| tags: |
| - continual-learning |
| - cpu-only |
| - no-gpu |
| - memory |
| - retrieval |
| - research-prototype |
| --- |
| |
| # Ananke v0 |
|
|
| **A memory that learns by reading.** Research prototype by **AMEFORGE**. |
|
|
| *Try the **[live demo](https://huggingface.co/spaces/ameforge/ananke)** — ask it anything it has read. The engine and its memory are proprietary to AMEFORGE.* |
|
|
| ## What it is |
|
|
| Ananke is a proprietary architecture developed at AMEFORGE, aimed at a simple goal: |
| **a model that learns by reading, the way a person does, without GPU training.** |
|
|
| - **Reads once.** Text is read a single time; there are no training epochs. |
| - **CPU only.** No GPU, no gradient-based training, no pretrained language model. It starts empty. |
| - **Keeps learning.** New text can be added at any time; what it read before is kept, reinforced |
| when read again, and slowly faded when unused. |
| - **Answers by quoting.** It answers factual questions by retrieving and citing the passages it read. |
| - **Says "I don't know".** When it has not read anything relevant, it refuses instead of inventing. |
|
|
| The demo memory was built by reading 18 English Wikipedia articles about West African countries |
| (about 165,000 words, 7,000 facts) in a couple of minutes on a standard CPU. |
|
|
| ## Evaluation (honest numbers) |
|
|
| Questions written **before** each evaluation and never used to tune the system: |
|
|
| | Question set | Right passage retrieved (top 5) | Exact short answer | |
| |---|---|---| |
| | Capitals, currencies, official languages, independence, largest cities (28 questions) | 18 / 28 (64 %) | 13 / 28 (46 %) | |
| | First presidents, colonial powers, neighbouring countries (27 questions, new relation types) | 20 / 27 (74 %) | 8 / 27 (30 %) | |
|
|
| What this means: |
|
|
| - **Finding the right passage works well**, including for relation types it was never tuned on. |
| - **Extracting the exact short answer from that passage is the main weakness**, especially for |
| answers that are lists ("which countries border…") or adjectives ("a *French* colony"). |
| - **Refusing on subjects it never read** (e.g. "What is the capital of Mars?") is a design priority; |
| it was not reliable in earlier versions and is being re-measured in the current one. |
|
|
| ## Limitations |
|
|
| - **Not a chatbot or an LLM.** It does not reason, combine facts or write free text. |
| - It **believes what it reads**: wrong text in, wrong answer out. |
| - It links two different words for the same idea only if its reading has connected them. |
| - English questions work best. Pronouns are understood only within the article about the subject. |
| - The short answer can be wrong even when the quoted passage is right: **always read the quotes.** |
|
|
| ## Intended use |
|
|
| Research and demonstration of learning-by-reading without GPU training. A self-updating, citable |
| memory over a document collection (encyclopedic text, documentation, notes). |
| Not for medical, legal, financial or safety-critical decisions. |
|
|
| ## Weights and code |
|
|
| Ananke has no neural weights to download. The engine and its memory are proprietary to AMEFORGE and |
| are not distributed; the Space provides public access for testing. |
|
|
| ## Contact |
|
|
| AMEFORGE — Mike Amega · [huggingface.co/AMFORGE](https://huggingface.co/AMFORGE) |
|
|
