Crate meme 

meme

Long-term memory for AI agents.

meme is a production-grade memory pipeline written in Rust. It persists knowledge extracted from conversations (or raw facts) to disk via LanceDB and tracks every change in a SQLite audit log.

Pipeline

1. Semantic Structured Compression — dialogues are windowed and sent to an LLM which extracts atomic, self-contained Memory entries with structured metadata (timestamps, locations, persons, keywords, …).
2. Lifecycle Reconciliation — each new entry is compared against existing memories in a single LLM call that decides ADD / UPDATE / DELETE / NOOP, preventing duplicates and resolving contradictions.
3. Intent-Aware Hybrid Retrieval — queries are analyzed by an LLM to produce a retrieval plan that drives parallel semantic (ANN), lexical (FTS), and structured-metadata searches. Optional reflection rounds iteratively refine coverage.

Quick Start

use meme::{Dialogue, Meme};

let meme = Meme::builder()
    .api_key("sk-...")
    .model("gpt-4.1-mini")
    .build()
    .await?;

// Ingest a conversation
meme.add(&[
    Dialogue::new("Alice", "Let's meet at 2pm tomorrow"),
    Dialogue::new("Bob", "Sure, see you at Shibuya station"),
]).await?;
meme.flush().await?;

// Store a fact directly (bypasses dialogue windowing)
meme.put("Alice prefers coffee over tea").await?;

// Hybrid search & Q&A
let memories = meme.search("Alice meeting").await?;
let answer  = meme.ask("When will Alice meet?").await?;

// CRUD
let all = meme.list().await?;
meme.update(all[0].id, "corrected content").await?;
meme.delete(all[0].id).await?;

// Audit trail
let events = meme.history(all[0].id).await?;

Public API

Entry point	Purpose
MemeBuilder	Fluent builder — configure API key, model, storage path
Meme	Runtime facade — add, flush, put, search, ask, CRUD, consolidate
Memory	A single self-contained unit of knowledge
Dialogue	Speaker + content input for conversation ingestion
Event / EventType	Change-history audit records
ConsolidationParams	Parameters for Meme::consolidate
ConsolidationStats	Summary returned by Meme::consolidate

Crate Layout

Module	Visibility	Contents
config	pub	Pure data configuration structs with validation
error	pub	MemeError enum and Result alias
model	pub	Domain types (Memory, Dialogue, Event, …)
store	pub	LanceDB vector store, SQLite history store, consolidation
embedding	pub(crate)	API and optional ONNX embedding providers
llm	pub(crate)	OpenAI-compatible LLM client, prompts, JSON schemas
pipeline	pub(crate)	Extractor, reconciler, hybrid retriever, answer generator

Re-exports

pub use error::MemeError;

pub use model::Dialogue;

pub use model::Event;

pub use model::EventType;

pub use model::Memory;

pub use store::ConsolidationParams;

pub use store::ConsolidationStats;

Modules

config

Configuration — pure data structures with defaults and validation.

error

Unified error types for the meme library.

model

Domain models — core data types for the memory system.

store

Storage layer — LanceDB vector store with multi-view indexing and SQLite history tracking.

Structs

Meme

The main entry point for the meme memory system.

MemeBuilder

Fluent builder for constructing a Meme instance.