Read the whole book from any address. One logical memory, forward-compatible across classical, photonic, and quantum hardware.
Today a language model generates causally — left to right — re-attending over an ever-growing transcript and re-deriving what it already knows. Aleph is a memory model in which the entire structure of an artifact is recoverable from any single point (the defining property of a hologram), so a generator can read the rest of the book instead of recomputing it. In the reference benchmark that turns an O(N²) access pattern into O(N) — a speedup that grows with size: 17× at 10 sections, 77× at 40, 145× at 100.
The same logical model runs on a hash table today and on a rare-earth quantum-memory crystal tomorrow, because the substrate is hidden behind a five-operation interface.
from aleph import AlephMemory, Codex, Slot
from aleph.backends import ClassicalBackend, PhotonicHolographicBackend, AFCBackend
mem = AlephMemory(ClassicalBackend(), codex) # runs today
# mem = AlephMemory(PhotonicHolographicBackend()) # optical, when available
# mem = AlephMemory(AFCBackend(finesse=12)) # quantum, when available — same code
view = mem.read_whole_book("section_7") # outline + constraints + recalled context, not the raw transcriptA hologram: every fragment encodes the whole. Spectral–spatial holography in rare-earth crystals is literally this, in frequency–space. Aleph lifts it to the logical layer: store information so the global structure is recoverable from any address. That is exactly the access pattern an LLM needs to stop re-deriving context.
Because one material family does all three. Rare-earth-ion-doped crystals (Eu³⁺:Y₂SiO₅, Pr³⁺:Y₂SiO₅, Er³⁺ hosts) store information:
| Regime | Mechanism | Demonstrated |
|---|---|---|
| classical | spectral hole burning (frequency-multiplexed bits) | decades of optical data storage |
| photonic | spectral–spatial holography / photon echoes | many patterns multiplexed in one volume |
| quantum | atomic frequency comb (AFC), EIT, CRIB/GEM | 1-hour optical storage; ~370-min coherence; 30+ h spin lifetime (2025) |
So "forward-compatible across substrates" isn't a bet on three technologies converging — it's three regimes of the same crystal. Room-temperature engineering uses phase-change (Ge–Sb–Te) and integrated photonics now; REIC is the quantum horizon. Full detail: docs/01-scientific-foundations.md.
python -m aleph.bench (numbers produced by the run, not hard-coded):
| slots | causal work | aleph work | speedup | recompute avoided | retries avoided |
|---|---|---|---|---|---|
| 10 | 4,195 | 244 | 17.2× | 30 | 5 |
| 40 | 92,560 | 1,204 | 76.9× | 120 | 35 |
| 100 | 610,030 | 4,204 | 145.1× | 300 | 95 |
Three savings, all real: bounded global view instead of a growing transcript (O(N²)→O(N)), derive-once content-addressed shared facts, and zero constraint backtracking because the whole plan is known up front. The benchmark's LLM is a faithful cost model, not a transformer — see docs/04 for exactly what is and isn't claimed.
The Aleph Inference Fabric (docs/07) doesn't replace the
proven techniques — it composes them on one content-addressed substrate. Each is a tier:
PagedAttention/vLLM (KV pages), prompt caching (prefix), RAG over HNSW/FAISS (long-term),
MoE-style routing, Merkle/IPFS dedup, LSM tiering — plus Aleph's novel "whole-book" working set.
Run it: python -m aleph.fabric → on a realistic workload it measures 97.5% prefix-hit and
2.47× KV-page dedup, every request served from a tier (real cache metrics, computed from the run).
Two honest measurements on live models:
- A single-call test (OmniCoder-9B via Ollama) measured ~1.0× — because independent-call token counting cannot see cross-request KV reuse, which is where the savings live.
- The served-stack test (
python -m aleph.kv_harness, llama.cpp:8774) measured it correctly: content-addressed prefix reuse cut prompt-eval time 58× and wall-clock 14× on a prefix-sharing workload. The token count is unchanged (the tell that misled the single-call test); the compute collapses because the shared prefix's KV isn't recomputed.
So: the serving-layer win is real and measured; the aleph.bench 17–145× figures remain
labelled as cost-model access-pattern properties, not transformer wall-clock. Full write-up,
diagnosis, and scope in VALIDATION.md.
application / any language ──────────────┐
│ spec/aleph.idl (5 ops + 1 address rule)
┌──────────────────────────────────────────┘
│ AlephMemory · Codex ("the book") · read_whole_book()
│
│ Backend interface: write · read · query · reconstruct · stats
├── ClassicalBackend RAM / phase-change (runs — reference)
├── PhotonicHolographicBackend spectral-spatial optics (simulation)
└── AFCBackend rare-earth quantum AFC (simulation)
Content addressing (BLAKE2b-128 of canonical payload) is the bridge: equal meaning → equal address, on every substrate, in every language.
git clone https://github.com/cognis-digital/aleph-memory && cd aleph-memory
python -m aleph.bench # the efficiency benchmark
python examples/demo_structured_generation.py # same loop, 3 substrates
pip install -e ".[dev]" && pytest -q # 10 testsPure standard library — no dependencies — so it runs anywhere.
aleph/— runnable reference (model, codex, llm cost-model, 3 backends, benchmark)docs/— the whitepaper: foundations · logical model · forward compatibility · read the whole book · spec & bindings · roadmap & open problems · inference fabric · glossaryVALIDATION.md— honest real-model validation (and why it measured ~1.0×)spec/— language-neutral interface + container formatbindings/— C ABI + Rust/Go/TypeScript surfaces (one interface, every language)
The classical backend, the Codex access model, the benchmark, and the tests run today. The
photonic and quantum backends are honestly-labelled physics-inspired simulations that prove the
interface is substrate-neutral and let you build against future hardware now. No speedup in this
repo is asserted without python -m aleph.bench reproducing it. The hard open problems are in
docs/06.
- One-hour coherent optical storage in an AFC memory (Eu³⁺:Y₂SiO₅) — Nature Communications (2021).
- Long optical coherence via dynamical decoupling in Eu³⁺:Y₂SiO₅ (~6 h class).
- Long spin lifetimes in Eu³⁺:Y₂O₃ ceramics for quantum memories — Communications Physics (2025).
- Time-bin qubit storage in rare-earth crystals — npj Quantum Information (2022).
This README synthesizes published results; it does not reproduce them.
COCL (Cognis Open Collaboration License) © Cognis Digital.