klieo-memory-pgvector

PostgreSQL + pgvector implementation of klieo-core's LongTermMemory (and FilterableLongTermMemory). Sibling to klieo-memory-qdrant — same trait surface, different store. Targets Aurora PostgreSQL Serverless v2 but works against any Postgres with the vector extension.

Schema

One table per deployment (<prefix>_v1, default klieo_facts_v1):

Indexes: HNSW on embedding (vector_cosine_ops) for ANN recall, and (scope_kind, scope_value) for the scope filter. Provisioned lazily on the first remember (so process starts don't race on DDL); the embedding dimension is fixed at first write.

metadata is jsonb (not a string) so relational joins from domain keys (Versicherungsnummer, Vertrag, GeVo-Typ) to facts remain possible against the same table — that join layer is a separate component, intentionally out of this crate.

Usage

# async fn example() {
use klieo_memory_pgvector::{MemoryPgvector, PgvectorConfig, DummyEmbedder};
use std::sync::Arc;

// Capability-shaped default (DummyEmbedder):
let mem = MemoryPgvector::connect("postgres://localhost/klieo").await.unwrap();

// Real embedder + config:
let cfg = PgvectorConfig::new("postgres://aurora.internal/klieo?sslmode=require")
    .with_table_prefix("triage")
    .with_embedder_id("titan-v2");
let mem = MemoryPgvector::new(cfg, Arc::new(DummyEmbedder)).await.unwrap();
let _ = mem.long_term;
# }

TLS / credentials

Carried by the connection URL (?sslmode=require), standard libpq behaviour — not a separate config knob. The table prefix is validated as a lowercase SQL identifier before it is interpolated into DDL.

Permissions

The first remember runs CREATE EXTENSION IF NOT EXISTS vector — the role needs permission to create the extension (on Aurora, rds_superuser), or an operator must pre-create it.

Embeddings

Model-agnostic via the shared Embedder trait. Pair with a Bedrock Titan Embedder on the AWS estate; DummyEmbedder (zero-vector, FIFO recall) is the default for tests/dev.