Crate memoir_core 

Memoir memory substrate as an embeddable Rust library.

Memoir stores what an agent is told and derives durable facts from it. A write is an episodic memory (the raw utterance); a background worker runs LLM extraction over it to produce semantic memories (the facts). See memory::MemoryKind. Recall reads the semantic layer; the episodic layer is the audit trail and the source the facts are re-derived from.

Two contracts govern how that derived knowledge is ranked and corrected.

Selection

Retrieval is not just cosine similarity. Each semantic memory carries a memory::Confidence (how sure the extractor was) and, once the categorize worker runs, a category label. client::Client::query blends these signals — cosine, confidence, recency, and a category bonus — via a client::RankingStrategy with tunable client::BlendWeights. Use client::Client::search for raw nearest-neighbor hits, query when you want ranked, prompt-shaped context. Hard filters (min_confidence, category) exclude rows; the blend weights them — distinct mechanisms.

Correction

Semantic memory is always derived, never hand-written — there is no API to edit a semantic row’s content. A consumer corrects a wrong fact by teaching: client::Client::feedback supplies the correction, and memoir re-derives from the episodic source. Editing the episodic source itself (client::Client::edit) cascades the same way. Both retire the stale derived rows and re-extract — see memory::RetirementReason for why a retirement is Rejected (a wrong extraction, counts against extraction accuracy) versus Stale (the source changed, does not). Retired rows are kept (not deleted) so accuracy is measurable via client::Client::extraction_stats.

Categorization

The category label is populated by a zero-shot NLI classifier (nli::NliClassifier), opt-in via nli::NliConfig on the client builder. Without it, categorization is skipped and the category-bonus blend term is inert.

Re-exports

pub use migration::DEFAULT_SCHEMA;

Modules

client

High-level facade composing the embedder, store, and vector index.

embedding

Text-to-vector embedding primitive.

graph

Property-graph abstraction for entity/relationship storage.

jobs

Postgres-backed write-behind queue for memoir-core’s worker.

llm

Text-completion abstraction for extraction and adjacent LLM tasks.

memory

Memory domain types.

migration

Memoir core’s database migrations.

nli

Zero-shot text classification via Natural Language Inference (NLI).

store

Postgres-backed source-of-truth storage for memories.

vector

Vector-index abstraction for similarity search.

Structs

Client

High-level facade composing the embedder, store, and vector index.

NliClassifier

Zero-shot text classifier using NLI entailment scoring.

ScoredLabel

A label paired with its entailment score from NliClassifier::classify.

Enums

ExecutionProvider

The hardware backend an NliClassifier runs its inference on.

NliConfig

Source of the zero-shot NLI classifier model.

NliError

Failure modes for the crate::nli::NliClassifier.

Constants

DEFAULT_EXTRACTION_PROMPT

System preamble for memoir-core’s extraction LLM call.

DEFAULT_SYSTEM_PROMPT

Default system-prompt section for memoir-core’s memory output.