Agent memory for AI agents

Short-term and long-term memory for AI agents: each conversation's recent messages and summary, and lasting facts extracted in the background and found again by meaning. Kept in your KiviDB...

An agent forgets everything between model calls unless something remembers for it. Agent memory keeps each conversation's recent messages and a running summary, and pulls the facts, preferences and events worth keeping out of conversations in the background, so the next conversation starts knowing them.

This conversation, and every one before

  • Working memory Each session keeps its recent messages and a running summary of the older ones, so a long conversation still fits in a prompt. It expires after the session goes quiet, an hour by default.
  • Long-term memory Facts and preferences (semantic) and things that happened (episodic), kept per user and found again by meaning, by keyword or both. Add up to three kinds of your own, each with a sentence saying what to keep.
  • Recall One call returns the summary, the recent messages and the memories most relevant to the user's latest message, as text ready to put before the model.

Extraction you do not write

  • Read once Once a session has been quiet for a few minutes, the store's language model reads the messages it has not read yet and decides what is worth remembering.
  • No duplicates A memory as similar as one the user already has is not kept twice.
  • Personal data removed By default emails, phone numbers, card numbers, US social security numbers and IP addresses are replaced with markers before a memory is kept.

Nothing leaves your account

  • Your keys Extraction runs on OpenAI, Anthropic or any OpenAI-compatible endpoint, and embeddings on OpenAI or a compatible endpoint, with your keys, stored encrypted. Your provider bills you for what they use.
  • Your database Sessions, memories and the search index live in the KiviDB database the store is in, under keys you can see. KiviDB charges nothing extra for agent memory.
  • From any agent Use the HTTP API, or the KiviDB MCP server's memory tools from Claude, ChatGPT and other MCP clients.

Questions

Which models can extract memories?

Any chat model on OpenAI or Anthropic, or behind an OpenAI-compatible endpoint. A small, fast model is enough: extraction reads a few thousand tokens of conversation and writes a few short sentences.

Can I keep my own kinds of memory?

Yes, up to three per store, each with a sentence describing what to extract, such as an order problem with its order number.

Does it work on sharded databases?

Not yet. A store's search index lives on one node, so agent memory runs on databases without shards.

Give your agent a memory

Create a database, add a memory store in the console's Agent memory section, and recall before every model call.