import type { Database } from "./database.js";
export declare class MemoryEmbeddingStore {
    private db;
    private insertStmt;
    private getStmt;
    private getAllStmt;
    private deleteStmt;
    constructor(db: Database);
    insert(memoryId: number, vector: Float32Array): void;
    get(memoryId: number): Float32Array | undefined;
    /**
     * Materialise the full embedding map. Eats memory linearly with the
     * memory count — only call this when you actually need every vector
     * (e.g. cluster-mode `prune`). Hot retrieval paths should use
     * `findTopK` / `getMany` instead.
     */
    getAll(): Map<number, Float32Array>;
    /**
     * Stream every embedding through cosine similarity against `query`
     * and return the top-K matches. Constant-memory in K rather than
     * linear in the table size — this is what `recall` and `remember`
     * should use when they only need the highest-similarity rows.
     */
    findTopK(query: Float32Array, k: number, minScore?: number): Array<{
        memoryId: number;
        score: number;
    }>;
    /**
     * Fetch a specific subset of embeddings by memory id. Used by `prune`,
     * which already has a bounded candidate set and doesn't need to load
     * the entire table. Batched to keep `?` placeholder count under the
     * SQLite limit.
     */
    getMany(ids: number[]): Map<number, Float32Array>;
    delete(memoryId: number): void;
}