AI memory built for production

Statewave is an open-source memory runtime for AI agents. It records raw events as immutable episodes, compiles them into typed semantic and episodic memories with provenance, and returns ranked, token-bounded context bundles that LLM applications can drop straight into a prompt.

An AI memory runtime — sometimes called memory infrastructure — is the persistence layer between an LLM application and the long history of users, projects, and decisions it interacts with. Instead of bolting a vector database to a chat history, a memory runtime stores structured memories with confidence scores, validity windows, and provenance, then assembles them into deterministic context bundles for each new question.

Your application sends raw events (conversations, tool calls, decisions) to Statewave as episodes scoped to a subject — a user, account, agent, or project. A pluggable compiler turns those episodes into typed memories. When the agent needs to answer a new question, it asks Statewave for context; Statewave ranks the relevant memories and episodes by recency, similarity, kind priority, and temporal validity, then returns a ranked, token-bounded bundle ready for the LLM.

A vector database stores embeddings and returns nearest neighbours. Statewave is a memory runtime that uses a vector store (pgvector) underneath, but adds typed semantic and episodic memories, deterministic ranking with explainable signals, idempotent compilation, provenance back to source episodes, temporal validity windows, conflict resolution, and a token-bounded context-assembly step. It also persists raw episodes append-only so memory can be recompiled.

Episodic memories capture what happened — append-only event records of a conversation, tool call, or decision at a specific time. Semantic memories are typed facts compiled from episodes — profile facts, preferences, procedures, and episode summaries — with confidence scores and validity windows. Statewave returns both in the same context bundle so the agent has long-term knowledge and the recent timeline.

Yes. Statewave is designed to be self-hosted. The storage layer is Postgres-only with the pgvector extension. A Docker Compose setup runs the full stack locally in two minutes; production deployments target Fly.io, Railway, or any container platform. There is no managed Statewave cloud — episodes and compiled memories stay in your infrastructure.

Yes. Statewave is open source under the Apache License 2.0 — server, Python and TypeScript SDKs, connectors, and this site. Apache-2.0 is permissive and includes an explicit patent grant, so you can use, modify, distribute, self-host, and build proprietary or commercial products with Statewave without a separate agreement. Optional enterprise support — SLA, indemnity, procurement, managed hosting, architecture review — is available as a services contract: contact licensing@statewave.ai.

Statewave is framework-neutral. Compilation goes through LiteLLM, which supports 100+ providers including OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex, Mistral, Groq, Together, and locally hosted models via Ollama or vLLM. The retrieval and context-assembly layer is provider-agnostic — your agent calls Statewave for context, then calls whatever LLM it likes. Python and TypeScript SDKs make integration with custom agents, copilots, and assistant frameworks straightforward.

Yes — anywhere your agent can make an HTTP call, it can call Statewave. The REST API and SDKs return ranked context bundles ready to splice into a system prompt, a tool result, or a retrieval-augmented chain. There is nothing framework-specific to install on the agent side; you decide where the context goes in your prompt.

Statewave runs as a Python service backed by Postgres with the pgvector extension. The default deployment is a single container plus a Postgres instance, orchestrated via Docker Compose. OpenTelemetry instrumentation is optional. Embeddings can be hosted (OpenAI, Cohere) or fully local (Ollama, sentence-transformers); the heuristic compiler runs entirely without an LLM if you want a fully local pipeline.