A model that learns continuously needs a principled way to decide what to keep. Keep everything and it drowns in noise: every typo, every one-off request and every half-finished thought becomes permanent. Keep too little and it forgets what matters. Human memory walks this line every day, consolidating what's important while you sleep and letting the rest fade. Living Memory is our attempt to give a machine the same judgement.

The problem with all-or-nothing memory

Most AI memory today is bolted on. A system saves snippets of past conversations to a database and stuffs them back into the prompt later. It works, up to a point, but the model itself never changes, and the memory has no sense of what's important. Ten trivial facts and one critical one are treated the same until the context window fills up.

Voltaic learns in its weights, which is far more powerful but raises the stakes. If every interaction can change the network, something has to decide which changes become permanent. That something is Hierarchical Synaptic Consolidation.

Three signals per synapse

HSC v2 tracks three signals for every synapse in the network:

  • Importance: how much the synapse contributes to behaviour you rely on. A connection that sits underneath a skill you use every day scores high.
  • Velocity: how fast it is changing right now. A synapse that's swinging wildly is still being learned and shouldn't be frozen yet.
  • Clarity: how consistent its signal is over time. A connection that keeps pointing the same way is trustworthy; one that flips back and forth is noise.

Synapses that score high across all three are hardened and protected from being overwritten. Those that don't stay free to adapt. Persistence has to be earned.

Why "hierarchical"

Consolidation doesn't happen only at the level of single connections. Skills are built from groups of synapses, and groups from larger structures. HSC applies the same logic at each level, so the network can protect a whole skill while still letting its fine details improve. This is also what lets HULL, our stability-hull model, work: HSC decides what to harden, and HULL keeps new learning from crossing into mastered territory. Together they took forgetting on CORe50-NI from 1.20% to 0.04%, as we describe in our benchmarks write-up.

MemoryForge: consolidation you can use

HSC is the mechanism. MemoryForge turns it into a working memory system that keeps what's important, merges what's related and lets noise fade over time.

  • Keeps what's important. Facts you rely on, preferences you've stated and decisions you've made are reinforced each time they prove useful.
  • Merges what's related. Three conversations about the same project become one coherent picture rather than three fragments.
  • Lets noise fade. One-off details that never come up again lose weight gradually, the way a phone number you dialled once fades from your own memory.

ContextFlow: the right memory at the right moment

Memory is only useful if the right part of it shows up when you need it. ContextFlow works out what's relevant right now, resolves conflicting information, and assembles exactly the slice of memory each query needs. If you told Halo last month that a meeting moved to Thursday and then moved it again to Friday, ContextFlow is what makes sure the answer is Friday, and can tell you why.

Memory you can see

A learning system you can't inspect is a liability. So what Voltaic learns is visible in a memory you can search, pin, edit and forget.

  • Search it to see what the system believes about a person, a project or a topic.
  • Pin things that must never fade.
  • Edit anything it has wrong.
  • Forget anything you'd rather it didn't keep.

And all of it is scoped to your workspace. Your memory adapts your model, for you only; it's never pooled into a shared dataset. We explain exactly what that means in Will my data be used to train the model?

Not a black box

The point of Living Memory isn't just to remember more. It's to stop you having to re-explain yourself to your tools every morning, while keeping you in charge of what they remember. That combination, a memory that improves with use and a memory you can audit, is what makes continual learning safe to deploy.