SRMU

A memory that stays useful as observations arrive

The Sequential Relevance Memory Unit (SRMU) is an update rule for high-dimensional vector memories. It reduces the influence of repeated information while allowing the memory to adapt when observations change.

An example: remembering device health

Imagine monitoring several devices. One reports its health often; another is observed only occasionally. The goal is to remember the current state of all devices, even though each new observation describes just one.

A vector memory stores each device’s identity and health as an association in a shared vector. Adding every report equally can let a frequently observed device dominate the memory. If that device’s health changes, its old reports also remain mixed in.

SRMU compares a new report with the value already represented for that device. A close match receives a small update; a different value receives a larger one. Decay gradually weakens previous contributions so the memory can follow changes over time.

How the update works

In vector symbolic architectures, a key identifies an entity, such as a device, and a value describes its state. Binding combines them into an association; adding associations builds a shared memory. Querying with a key retrieves an approximate value.

From observation to comparison to updated memory

Inside the memory unit

The full diagram shows how these paths meet. The persistence parameter γ controls decay: smaller values retain less of the past, while γ = 1 preserves the prior memory. The relevance weight w = 1 − s controls how strongly the incoming association is added, where s is its similarity to the retrieved value.

The two controls serve different purposes: decay reduces the influence of the past, while the relevance gate limits reinforcement of observations that already match the memory. The gate acts when information is stored, so it can complement methods that clean up estimates during retrieval.

What the experiments show

The strongest result comes from the experiment combining uneven sampling and changing states. At the same decay setting (γ = 0.95), SRMU increased similarity to the current ground-truth memory from 0.755 to 0.850: a 12.6% relative improvement over decay alone.

What was tested

The synthetic monitoring task uses five devices, five possible health states, and 256-dimensional vectors. Only one device is observed per timestep, with a 5% chance of observation noise. Three experiments isolate uneven sampling, gradual state changes, and their combination with occasional abrupt changes. The plots report 1,000 trials.

Cosine similarity measures how closely the learned memory aligns with a ground-truth vector encoding the current state of every device. Higher is better. Memory-vector magnitude measures the norm of the accumulated vector and is interpreted alongside similarity.

Combined-experiment comparison

Source: Table 2 in the paper. SRMU with decay also reduces vector magnitude by 53.5% compared with decay alone. Together with better alignment, this indicates less redundant accumulation. All methods use the same vector dimensionality; this result does not measure a reduction in storage bytes.

The other experiments clarify when each control helps. With uneven sampling and fixed states, SRMU without decay has the highest similarity (0.942). When states drift, SRMU with decay performs best (0.907). Decay is useful when the state changes, but can weaken useful information when it remains valid.

Scope and limitations

The evidence comes from a small synthetic state-tracking task. Larger sensing systems, spatial maps, and neuromorphic hardware remain to be evaluated. The persistence parameter still sets a tradeoff: stronger decay removes old contributions faster but can also weaken useful associations that are rarely observed.

SRMU controls how observations enter memory; it does not guarantee that a new observation is accurate or that retrieval is exact. Formal capacity and interference analysis, larger-scale evaluations, and efficient neuromorphic implementations are directions identified in the paper.

Connection to spatial memory

VSA-OGM studies how to encode spatial observations in vector memory and query them to reconstruct a map. SRMU suggests a next question: can that memory stay representative when some places are observed repeatedly and others change between visits?

In a spatial application, a key could represent a location and a value could describe occupancy. Repeatedly seeing the same occupied location should add less reinforcement, while observing a change at that location should have more influence. Decay would reduce the persistence of earlier observations.

The next step is to test whether this improves map updates while preserving useful information about rarely visited places. This is a proposed application; the published SRMU experiments evaluate device-state tracking.

My contribution

I developed the idea, designed and implemented the method, ran the experiments, analyzed the results, and wrote the paper. The paper was coauthored with Andrew Capodieci, David Gorsich, and Maryam Parsa and published at ICONS 2026.

@inproceedings{snyder2026srmu,
  title     = {SRMU: Relevance-Gated Updates for Streaming Hyperdimensional Memories},
  author    = {Snyder, Shay and Capodieci, Andrew and Gorsich, David and Parsa, Maryam},
  booktitle = {International Conference on Neuromorphic Systems (ICONS)},
  year      = {2026},
  doi       = {10.1145/3822454.3822487}
}