FHRR Holographic Memory

What if AI
never forgot?

Not a longer context window. Not a better search index. Native long-term memory — algebraic, content-addressable, and sub-millisecond at 30 Hz.

0
Cleanup Accuracy: <132 μs
p99 Recall Latency: 0
Codebook Concepts: 0
Dimensions (D)

How It Works See comparisons →

The Problem

Every AI forgets. By design.

Context windows are scratch pads. RAG is a prosthetic. Fine-tuning overwrites the past. None of them are memory.

Context Windows

Everything re-read, re-attended, re-processed every call. Quadratic cost. Signal degrades as relevant facts sink beneath noise — the "lost in the middle" problem.

Working scratch pad, not memory

RAG Retrieval

Retrieves text chunks by embedding similarity. Cannot bind structured relationships. No partial-cue recall. No temporal continuity. 100ms–1s latency per query.

External prosthetic, not native memory

Fine-Tuning

Updates weights to encode new facts, but catastrophic forgetting degrades old knowledge. Offline, batch-oriented, expensive. Cannot happen at conversation speed.

Weight overwrite, not associative recall

The Solution

Holographic memory in three operations

Bind, recall, cleanup. Algebraic operations in Fourier space — not similarity search.

Bind

Element-wise complex multiplication in FFT domain. Role-filler associations superpose in a single D-dimensional vector.

F_trace = F_key ⊙ F_value

Recall

Phase-conjugate unbinding: one multiply to recall. Probe from Kuramoto state, conjugate multiply against stored trace.

F_recalled = F_trace ⊙ conj(F_probe)

Cleanup

Correlate noisy recalled vector against the 744-entry codebook. Snap to nearest concept. Partial cues work by design.

k* = argmax cosine(codebook, recalled)

Persist

WAL writer, ETS hot tables, Mnesia metadata, Postgres audit trail. Merkle BLAKE3 integrity chain per recall.

Crash-safe holographic traces

Capacity

99.97% Cleanup Accuracy

At N=744 bindings, D=16,384 dimensions. Partial cues with as little as 30% overlap still reconstruct the correct concept via phase-coherent recovery.

bounded cleanup accuracy at capacity

Compression

qFHRR: 16× Compression

Quantized phase FHRR uses discrete cyclic group Z_K with K=16 phase bins. Integer-only arithmetic. 128 KB → 8 KB per vector. 0.987 fidelity retained.

16×

memory compression, L1 cache footprint

Memory Palace

Seven rooms, seven domains

Memory is partitioned by domain. Each room is a GenServer under a DynamicSupervisor — crash one, others survive.

:security

Audit logs, access rules, threat models

:agents

Transcripts, skills, session context

:build

Specs, CI outputs, validation evidence

:trinity

White paper, monograph, architecture

:infra

Deployment configs, mesh topology

:neuromorphic

NIR exports, spiking weights

:meta

Coherence aggregates, introspection

Head to Head

Memory approaches compared

Dimension	Trinity FHRR	Context Window	RAG	Fine-Tuning
Persistence	✓ WAL + Merkle	Session only	Index-dependent	Weight overwrite
Recall Mechanism	Phase-conjugate	Re-attention	Embedding search	Forward pass
Compositional	✓ Role-filler binding	Implicit	✗ None	Implicit
Partial-Cue Recall	✓ Native cleanup	✗	~ Approximate	✗
Latency @ 30 Hz	<132 μs	Prohibitive	100ms–1s	Offline
Inspectability	✓ Merkle + coherence	Transcript	Chunks	Opaque weights
LLM Coupling	Orthogonal substrate	Same model	External	Same weights

The LLM forgets because it was never given memory. Trinity gives it a palace.

Seven rooms. 16,384 dimensions. One conjugate multiply to recall.

Explore Technology See Benchmarks Request Investor Deck →

What if AInever forgot?