Vector-based retrieval via Qdrant with scoped filters for userId, agentId, and runId.
Extracts atomic facts from content before embedding to ensure richer and more accurate recall.
Tracks agent execution history with step-by-step action logging and auto-generated summaries.
Retrieves relevant memories, optionally reranks them, and generates context-aware answers using LLMs.
Uses normalized content hashing to efficiently prevent duplicate memory entries.
Ensures AI data consistency with robust Zod schemas for all memory inputs and queries.
A clean, secure gateway for your applications to store, search, and retrieve memories. Simple endpoints that validate every request and return structured responses.
The brain of the system. Coordinates how memories are created, prevents duplicates, extracts key facts, and orchestrates the entire storage and retrieval flow.
Converts your content into searchable vectors using AI models. Enables semantic understanding so you can find memories by meaning, not just keywords.
Dual storage for reliability and speed. Structured data lives in a relational database while vectors are stored in a specialized search engine for instant retrieval.
The system supports semantic memory for knowledge storage, episodic memory for event history, and procedural memory for tracking agent execution steps and action logs.
Content is processed through an LLM to extract atomic facts before embedding. This ensures each memory chunk contains discrete, searchable knowledge rather than raw text passages.
We use Qdrant as the primary vector store, with support for scoped filtering by userId, agentId, and runId to ensure proper memory isolation across sessions.
Content is normalized and hashed before storage. Duplicate entries are detected and prevented at the ingestion layer, keeping your memory store clean and efficient.
Yes, the memory layer is framework-agnostic and exposes a simple API. It works with LangChain, AutoGPT, custom agents, and any system that can make HTTP requests.
All memory entries are scoped to specific user and agent contexts. Data is never shared across tenants, and you can self-host the entire stack for complete control.