User Guide

Persisting provides unified tiered storage for AI workloads. Trajectories, parameters, and KV cache share the same addressing model (TTAS), storage engine (Lance), and distribution runtime (Pulsing).

Core: Unified Storage

Guide	Workload	Dimensions	Status
Tensor Memory	Parameters, KV Cache, Trajectories	(param_id, shard), (session, layer, head, time), (run_id, time)	🧪 Experimental

All three workloads use persisting.open() with the same TTAS addressing. Block-tiered across host memory and SSD (GPU planned).

Tools on the Same Foundation

Guide	Description	Status
Capture	Proxy and record LLM traffic — persisting traj	✅ Stable
Queue	Append/consume event streams, KV API, samplers	✅ Stable
Search	Document indexing and vector/hybrid search	✅ Stable
Compute	Map-style task orchestration — plan() + execute()	✅ Stable
Custom Backends	Implement your own storage backend	📖 Reference

---

Architecture

┌───────────────────────────────────────────────────────┐
│  Trajectories      Parameters        KV Cache        │
│  (run_id, time)    (param_id, shard) (sess, layer, …)│
├───────────────────────────────────────────────────────┤
│                    TTAS                                │
│            Tiered Tensor Address Space                 │
├───────────────────────────────────────────────────────┤
│   Tiering:  GPU (L0)  ↔  Host (L1)  ↔  SSD (L3)     │
│   Route:    Pulsing actor runtime                      │
├───────────────────────────────────────────────────────┤
│            Lance Columnar Storage                       │
└───────────────────────────────────────────────────────┘

Choosing Your Entry Point

You want to…	Start with
Store/retrieve parameters or KV cache by tensor subscript	Tensor Memory
Record agent LLM calls	Capture
Stream events with persistence	Queue
Index and search documents	Search
Run batch jobs with checkpoint/resume	Compute
Plug in custom storage	Custom Backends