Overview

Aletheia Memory Engine Overview

Aletheia is the purpose-built memory infrastructure for AI agents and applications that need to understand time, state, and evolving user truth—moving far beyond the limitations of a flat vector index.

What Aletheia is built to do

Building AI agents that remember users over long periods of time is hard. Standard RAG (Retrieval-Augmented Generation) approaches dump raw chat logs into vector databases, which quickly leads to bloated context windows, contradictory facts, and terrible performance on counting or math queries. Aletheia solves this by acting as an intelligent memory controller.

Adaptive Hybrid Retrieval: We fuse semantic search with lexical search and neural reranking, dynamically shifting weights based on whether you are asking for a concept or a specific number.
Temporal Awareness: Aletheia keeps fresh context visible through time-aware ranking algorithms, ensuring old memories decay gracefully.
Fact Supersession: We track when facts change (e.g., moving to a new city) so stale claims stop competing with the updated truth in your LLM's context window.
Deterministic Aggregation: Counting and math queries are resolved deterministically before they reach the LLM, ensuring perfect accuracy.
Seamless Developer Experience: Run the exact same Rust-powered engine locally as a sidecar during development, and deploy it to the cloud for production.

Why engineering teams choose Aletheia

Most memory systems on the market can retrieve chunks of text. Very few can update beliefs cleanly, handle mixed factual and episodic recall, or ship a local-first workflow that matches production performance.

Aletheia is designed to fill that exact gap. It distills noisy chat logs into clean, queryable facts, saving you tokens and preventing your AI from hallucinating about past interactions.

Core design principles

1. Hybrid retrieval is mandatory

Approximate Nearest Neighbor (ANN) search alone is not enough for human-like memory. Aletheia combines semantic search, BM25-style lexical search, and cross-encoder reranking so both exact phrases and latent meaning are respected.

2. Not all memories are equal (Memory Classes)

A passing thought shouldn't be ranked the same way as a core user preference. Aletheia categorizes memories into classes (Facts, Summaries, Preferences, Episodic traces) and applies different Time-To-Live (TTL) and decay policies to each.

3. Local-first development accelerates shipping

You shouldn't need a cloud API key just to run your test suite. Developers can run the Aletheia Rust engine locally as a lightweight binary before pointing their Python or Node.js SDK at a hosted environment.

Repositories

Explore our public repositories for the core platform, SDKs, and model adapter surfaces:

Core engine repository is private: Aletheia

Fundamentals