Local-first TypeScript SDK

Shared semantic memory for AI agents

agentOrc gives multiple agents one persistent memory they can all write to and search — stored locally in SQLite, retrieved by meaning, with a tiny TypeScript API.

No Redis glue. No JSON scratch files. No custom locking. Just remember() and recall().

$ npm install agentorc

Planner writes · worker recalls · same shared memory

Why it exists

Agents forget. Shared state gets messy.

Most multi-agent setups store knowledge in process memory, JSON files, or a generic key-value store. That works until you need persistence, concurrency, and search by meaning — not just by key.

Without agentOrc

  • Each agent keeps its own isolated context
  • Developers invent shared state with globals or Redis
  • JSON dumps become the source of truth
  • No semantic search across what agents learned
  • Concurrency and crash safety are DIY

With agentOrc

  • One shared semantic memory for the whole system
  • SQLite persistence with WAL and ACID writes
  • Natural-language recall via embeddings
  • A small, stable TypeScript API
  • Runs locally with zero infrastructure

How it works

Write once. Recall by meaning.

The whole product fits in one loop: store text as vectors, keep the original, search later with natural language.

  1. Step 1

    Agent remembers

    One agent calls remember() with a fact and optional metadata.

  2. Step 2

    Embedding generated

    content.text is embedded through your OpenAI-compatible endpoint.

  3. Step 3

    Stored in SQLite + sqlite-vec

    Original text, metadata, and vector land in a local ACID write.

  4. Step 4

    Another agent recalls

    A different agent asks in natural language and gets ranked hits.

Core features

Built for real agent systems

Local-first

Everything lives in a SQLite file on disk. No hosted vector DB, no cloud dependency for the memory layer.

Semantic recall

Agents ask in natural language. agentOrc embeds the query and returns the closest memories by meaning.

Multi-agent by design

Many agents can write and read the same store safely. WAL mode and transactions handle concurrency.

SQLite + sqlite-vec

Durable storage with an in-process vector index. Back it up like any other database file.

Tiny API

remember, recall, compress, forget, history, stats. Easy to drop into an existing agent loop.

Zero infrastructure

No Redis cluster, no queue, no sidecar. Install the package and point it at a file path.

Install and go

One dependency. Point it at a SQLite path and an OpenAI-compatible embedding endpoint.

$ npm install agentorc

Read the docs

API reference, configuration, concurrency notes, and guides for shared multi-agent memory.

Open Documentation