pgmem docs

All tuning lives in one PgMemConfig you build once and pass as PgMem(dsn, config=...). Every field has a sensible default, so you only set what you care about. Flat PgMem(...) keyword overrides still work and win over the same field in config.

from pgmem import PgMem, PgMemConfig, PoolConfig, IngestionConfig

mem = await PgMem.create(
    dsn,
    embedder=embedder, llm=llm,
    config=PgMemConfig(
        pool=PoolConfig(max_size=20),
        ingestion=IngestionConfig(edge_dedup_cosine_threshold=0.97),
    ),
)

Providers (embedder / llm / cross_encoder) and the policies (claim_policy) are not in PgMemConfig — they're injected directly into PgMem, so config stays pure, loggable data with no secrets.

PgMemConfig

Field	Type	Default	When to touch
`pool`	`PoolConfig`	`PoolConfig()`	Concurrency / connection sizing.
`fulltext`	`FullTextConfig`	`FullTextConfig()`	Choosing a BM25 backend.
`ingestion`	`IngestionConfig`	`IngestionConfig()`	Tuning dedup, retries, size limits.
`search`	`SearchConfig \| None`	`None` (auto-pick)	Forcing a specific retrieval recipe.
`compiler`	`CompilerConfig`	`CompilerConfig()`	Retrieval depths, tokenizer, live-context provider.
`community`	`CommunityConfig`	`CommunityConfig()`	Community rebuild limits.

PoolConfig

Field	Default	Notes
`min_size`	`1`
`max_size`	`10`	Needs ≥ 2 for `add_episode_bulk`, `compact_async`, and concurrent application/worker activity. Size it for the groups you ingest concurrently. A poll-only projection worker does not hold an idle listener connection.

FullTextConfig

Field	Default	Notes
`backend`	`FullTextBackend.auto`	`auto` uses an installed BM25 extension if present, else native Postgres FTS.
`language`	`"english"`	Must match the regconfig baked into the generated `tsvector` columns (English today).

CompilerConfig

Controls context compilation.

Field	Default	Notes
`strategy_depths`	see below	Per-strategy retrieval depths. Every built-in strategy must be present.
`tokenizer`	`TokenizerConfig()`	Token accounting for budgets.
`injected_context_provider`	`None`	A `(session, query) -> list[str]` coroutine for always-injected Tier 6 context. Runs on the turn — keep it fast; prefer the per-call `injected_context=` argument for network calls.

Default strategy_depths (episodic / narrative / semantic):

Strategy	episodic	narrative	semantic
`conversation`	5	3	5
`task`	8	0	3
`recall`	3	8	8

RetrievalDepths

RetrievalDepths(episodic, narrative, semantic) — non-negative ints; 0 disables a tier. These are retrieval limits, not token reservations; the compiler still drops recalled material when the budget requires.

TokenizerConfig

TokenizerConfig(model=None, encoding=None) — set a model (if tiktoken knows it) or an encoding, not both. Both unset → the default o200k_base encoding.

IngestionConfig

Tuning and safety limits for add_episode. Commonly tuned:

Field	Default	Notes
`node_dedup_candidate_limit`	`15`	Cosine candidates considered when deduping an entity.
`node_dedup_min_score`	`0.6`	Similarity floor for an entity dedup candidate.
`edge_invalidation_candidate_limit`	`10`	Candidates checked for contradiction.
`edge_dedup_cosine_threshold`	`None`	Optional fast-path edge dedup by cosine.
`previous_episode_limit`	`10`	Prior episodes fed as extraction context.
`retries`	`3`	Embedder/LLM retry attempts.
`bulk_concurrency`	`20`	Parallel per-episode resolvers in bulk ingest.
`update_communities`	`False`	Incremental community update after each episode.
`expected_embed_dim`	`1536`	Must match the schema's `vector(N)` width.

Plus size guards (max_name_chars=512, max_fact_chars=8192, max_content_chars=65536, max_summary_chars=1000) that protect rows from oversized input — raise them deliberately.

CommunityConfig

Field	Default	Notes
`max_concurrent_builds`	`10`	Cap on simultaneous per-community LLM reductions — set to your provider's limit.
`label_propagation_max_iterations`	`100`	Clustering loop cap.
`max_summary_chars`	`1000`	Community summary cap.

Policies (injected, not in config)

Object	Passed as	Concept
`ClaimPolicy`	`PgMem(claim_policy=...)`	Trust & memory formation — governs claim formation, evidence, conflicts, and live-state routing.

PgMem — where config is passed
The context compiler — what CompilerConfig tunes
Tune retrieval — strategy depths and budgets in practice

Configuration