DeltaStore

Store changes, not state.

Most AI systems are quietly drowning in their own memory. Every snapshot of agent state, every reasoning trace, and every branching simulation path gets stored in full, over and over again. DeltaStore takes a fundamentally different approach: store changes, not state. The result is lower infrastructure cost, practical time-travel debugging, and auditability by design.

593x
Peak Compression
14
Domain Modules
2
Revenue Tracks
O(d)
Reconstruction Cost
Narrated Demo

IRIS presenting DeltaStore

Watch IRIS walk through how DeltaStore compresses state in real time while keeping every frame fully reconstructable.

Why It Matters

Auditability becomes a side effect of storage

Traditional AI systems repeatedly store full snapshots of state. DeltaStore records only the minimal change required to move from one state to the next, while still preserving full reconstruction.

For AI infrastructure Replay buffers, agent memory, reasoning traces, branching simulations, and time-travel debugging all become cheaper and more inspectable.
For healthcare and PBM pricing DrugPriceChain applies delta architecture to drug pricing itself: every markup, fee, rebate, clawback, and formulary change becomes a signed, timestamped audit trail.
For enterprise buyers The value is not just compression. It is verifiable history: what changed, when it changed, and how the system got there.

Production Compression

Convergence, time-travel debug, RL replay, deep branch trees, and cache layers all show order-of-magnitude storage reduction when you store deltas instead of full snapshots.

Business Model

Track 1 is healthcare and drug pricing auditability. Track 2 is DeltaStore as a drop-in memory layer for enterprise AI stacks.

Market Position

Vector databases store embeddings. Experiment tools track runs. Event systems are generic. DeltaStore targets domain-specific state compression with audit-ready reconstruction.

Temporal State Compression

TSC is DeltaStore applied to transformer memory

Temporal State Compression treats the token dimension of a transformer KV cache like a temporal signal: store keyframes, store deltas between them, and reconstruct exactly or approximately later. It is the same DeltaStore idea, specialized for model memory.

Paper-style result On the current GPT-2 KV benchmark path, TSC reaches 7.98x lossless and 63x lossy compression at sequence length 1024, while preserving identical top-1 token choices at the tested settings.
Live runtime result On the real model-consumed Hugging Face cache path, the current frontier is more conservative: roughly 1.04x live KV compression with exact top-1 behavior at the best safe operating point. That path is real, but still being tuned.
Why it matters Most teams introducing AI agents lack persistent, portable, governed memory. TSC gives DeltaStore a model-native way to compress that memory substrate without forcing companies to rebuild their stack around one provider.
Current Frontier

Two truths, one roadmap

63x
Offline KV Compression
kf=64 @ seq=1024
1.039x
Safe Live KV Frontier
exact top-1
  • The paper path shows that temporal compression is a real primitive for transformer memory, not just generic storage theory.
  • The live-cache path shows the harder product problem: selectively compress only the parts of model memory that are cheap to approximate.
  • That is why DeltaStore matters beyond one benchmark. It gives us a reusable temporal storage and replay fabric, and TSC becomes one specialized codec family inside it.
Interactive Demo

See compression happen in real time

Run the DeltaStore visualizer directly below and watch traditional full-frame storage diverge from delta-based storage as state evolves.

Two-Track Revenue

  • Healthcare vertical: DrugPriceChain for PBM pricing transparency and employer auditability.
  • AI infrastructure licensing: DeltaStore as a memory layer for companies paying large recurring storage costs today.
  • Same core primitive, two distribution motions.

Domain Modules

  • Healthcare formularies
  • RL replay buffers
  • Vehicle telemetry
  • Tax code versioning
  • Time-travel debugging
  • Perspective synthesis and adversarial evolution

DeltaStore turns memory from a linear cost into a compressed timeline.

IRIS DELTA ANALYST