Skip to Content
These docs describe PRME 0.16.0a8, an alpha. pip install prme installs the latest stable release. To use these docs as written, run pip install prme==0.16.0a8. Changelog
FeaturesMultiple local packs

Resource control for multiple local packs

Separate directories are currently the supported way to keep two named projects isolated. Use one pack per project, even when they share a user and Scope.PROJECT. A metadata label does not prevent entity matching or maintenance across projects.

Each open DuckDB database has its own worker pool. Applications opening several packs can explicitly choose a worker count:

from prme import MemoryEngine, config_from_directory
from prme.storage.embedding import FastEmbedProvider
 
provider = FastEmbedProvider()  # Caller-owned; share on the application's loop.
config = config_from_directory("./projects/aurora")
config.database_url = None
config.duckdb_threads = 1
 
async with MemoryEngine.open(config, embedding_provider=provider) as memory:
    await memory.store("Retention is seven days", user_id="alice")

PRME_DUCKDB_THREADS=1 configures the same setting through the environment. An explicit constructor argument takes precedence. The default is None, which preserves DuckDB’s default. This setting applies when opening a local engine; PostgreSQL ignores it. It controls DuckDB workers, not embedding inference, Tantivy, USearch, Python workers, total process memory or all native threads.

Concurrent engines opening the same file must use matching thread settings, including matching explicit/default choices. DuckDB rejects conflicting opens without changing the already-open engine. Close all connections before changing the setting for that file. Worker settings are runtime configuration, not durable memory content; reopening a copied pack uses the caller’s configuration.

Embedding inference has its own setting. embedding.threads (PRME_EMBEDDING__THREADS) sets the ONNX Runtime threads of a fastembed model when it loads, and a caller-owned FastEmbedProvider(threads=1) takes the same value. The default is None, which keeps ONNX Runtime’s choice of one thread per core the machine shows; that choice does not follow a container’s CPU limit. The setting is not recorded with a vector, and it changes how many threads compute each vector, not the model or the version recorded with it. On the machine measured when the setting was added, one thread gave the same vectors, bit for bit, as the default for 300 texts; check this on your own platform before changing it for a store whose vectors already exist.

Closing idle engines can reduce active resource use; applications should keep the engine context alive until all operations finish. A shared provider avoids loading another embedding model per pack. It does not share graph or search indexes. A bounded cache must also handle leases, concurrent open/close, cancellation and resource ownership. The local workspace API now provides that bounded cache with lease ownership and durable project identity.

The completed resource probe  uses tiny authored packs and a shared warmed model. Results do not establish an optimal thread count for larger histories or concurrent workloads. Measure your application before changing its worker count.

DuckDB describes its per-instance thread configuration in the configuration reference  and the runtime scope of settings in its SET documentation .

Source: docs/LOCAL-RESOURCES.md  in PRME at bd92166f.

WorkspacesCustom embeddings