API Reference

Persisting's APIs share a common core — TTAS addressing, Lance storage, Pulsing distribution.

Module	Use for	Status
Tensor Memory	persisting.open() — unified storage for parameters, KV cache, trajectories	🧪 Experimental
Queue	persisting.Queue — event streaming, KV interface, samplers	✅ Stable
Search	persisting.search — document indexing and retrieval	✅ Stable
Tensor Address Space	persisting.core — direct access to Dimension, Region, canonicalization	🧪 Experimental

---

Unified Tensor Storage

persisting.open(namespace, dims, ...) → TensorNamespace
kv[key]                              → Handler
h.tensor()                           → ndarray
h.put(data)                          → None

One interface for all three workloads:

Namespace	Dims	Use
params/llama-70b	(param_id, shard)	Model weights
kvcache/v1	(session, layer, head, time)	KV cache
trajectories/v1 (planned)	(run_id, time)	Trajectory tensor access

→ Tensor Memory API

Queue & Events

Queue(name, storage_path, ...)
q.put(record) / q.get(limit)
KVInterface(q).kv_put / kv_batch_get

→ Queue API

Search

add_document(dataset, text)
query(dataset, query, mode, k)

→ Search API

TTAS Types

Dimension(name, kind)
TensorView(dims)[key] → Region
canonicalize(region) / project_prefix(region, dims)

→ Tensor Address Space API