Symbol vectors
Each concept gets a unique 1024-dimensional vector, generated deterministically from its name. Unrelated symbols come out nearly orthogonal.
About this project
About HoloMem
How holographic memory works, for engineers who want the details.
What is holographic memory?
Holographic memory is a brain-inspired way to store and retrieve information. Instead of indexing memories by keys or keywords, each memory becomes a high-dimensional vector trace built from a few algebraic operations. To retrieve, you construct a probe vector and find the stored traces most similar to it. It’s an approximate, content-addressable lookup, not an exact database query.
HRR operations
Holographic Reduced Representations (HRR) use three core operations to pack structured knowledge into fixed-size vectors.
Binding (circular convolution)
Associates two concepts. bind(SUBJECT, “user”) produces a vector representing “the subject is user”. Implemented via FFT. The result looks dissimilar to both inputs.
Superposition (addition)
Multiple bindings are summed into a single trace vector. All associations live in one fixed-size vector. Lossy, but approximately recoverable.
Unbinding (circular correlation)
Given a trace and a key, recovers the bound value approximately. Cleanup memory maps the noisy result back to the nearest known symbol.
Retrieval modes
- Keyword: Token overlap scoring (baseline).
- Holographic: Cosine similarity between probe and trace vectors.
- Hybrid: Weighted combination of holographic 40%, keyword 30%, trust 15%, entity overlap 15%.
Tradeoffs vs vector DB / RAG
| Aspect | HRR | Vector DB + RAG |
|---|---|---|
| Embedding | Deterministic, no ML model | Requires embedding model |
| Structure | Encodes roles and relations | Flat semantic similarity |
| Scale | Prototype (hundreds) | Production (millions) |
| Dependencies | NumPy + SQLite | External services, API keys |
| Explainability | Component scores, matched symbols | Opaque similarity score |
Tech stack
Backend
FastAPI, SQLAlchemy, NumPy, SQLite
Frontend
Next.js, TypeScript, Tailwind, shadcn/ui
Math
FFT-based circular convolution, cosine similarity
Storage
SQLite with BLOB vectors, JSON metadata