HoloMem

Playground · HRR Lab and applied sandbox

Run the operations. Stress-test the memory.

Below are four HRR interactives (unbind, capacity, role-binding, noise), followed by a sandbox for the applied memory system from the homepage. Seed the lab to load example memories you can teach against, query, and disturb.

  1. 01

    Seed the lab

    Load a large set of example memories so there is something rich to query against.

  2. 02

    Teach or browse

    Add facts in plain English on the left, watch them join the memory field in the middle.

  3. 03

    Recall and disturb

    Ask a fuzzy question on the right. Below, inject noise or create contradictions and see how trust adapts.

HRR lab · what only this system can do

Four operations a keyword index can’t imitate

The recall scorer in the applied sandbox mixes HRR similarity with keyword overlap and trust. That makes HRR look like a noisier inverted index. These four interactives use only the operations HRR provides (binding, unbinding, superposition, cleanup) and show where the algebra actually earns its keep.

01

Unbind and cleanup

The headline HRR operation. Store a fact as a bound vector, then extract a value by 'dividing out' the role. Keyword search can re-rank stored strings. Only HRR can recover a bound value as a vector.

Plus 3 distractor facts superposed into the same trace (e.g. atlasusesfastapi).
03

Roles vs. bags of words

"Maya manages the auth service" and "The auth service manages Maya" contain identical tokens. Keyword scoring can’t tell them apart. Role-binding can.

Memory A

Maya manages the auth service.

Memory B

The auth service manages Maya.

02

Capacity in one vector

How many (key, value) pairs fit into a single 1024-d trace before cleanup starts losing them? Each point is the recall accuracy averaged over 3 trials.

04

Graceful degradation under noise

Add Gaussian noise to the encoded trace, then try to recover the OBJECT of one of the bound facts. HRR confidence drops smoothly. A keyword index has no analog: once your query loses its tokens, it returns nothing.

01Teach

Type a fact, preference, or observation. It gets encoded into a 1024-dim trace and joins the field on the right.

02Memory field

Memory field is empty

Click Seed the lab above, or teach a memory on the left. Nodes will appear here and cluster by shared entities.

force-directed graph

03Recall

Ask something fuzzy, partial, or indirect. Matching memories light up in the field, with each component score broken out.

Weighted blend: 40% vector, 30% keywords, 15% trust, 15% entities.

Recall duel

Run the same query through plain keyword search and holographic recall side by side. The differences show up most clearly on indirect or paraphrased questions.

Distortion lab

Stress-test the system. Add noise, create contradictions, watch how trust adapts. Each action updates the memory field above.

Inject noise

Add 3 random low-trust nonsense memories. Watch them appear as dim nodes in the field above.

Contradict

Pick a memory to contradict. The system creates a conflicting version and flags both.

No memories to contradict yet

Reset

Wipe all memories, vectors, and symbols. Start fresh.

How it works

Three operations power holographic memory. Click each step to walk through bind, superpose, and unbind with worked notation side by side.

Bind

Each fact is encoded as role-filler pairs. 'subject' is bound to 'Maya' by convolving their symbol vectors. This creates a single vector that represents the relationship.

subject ⊛ Maya → bound_vector

The binding is invertible: given the role vector, you can recover the filler.