Introduction

What can I build?

Real things you can build when your agent can remember, explained without jargon.

Aether is the memory layer for AI agents: remember what's worth keeping, recall the right context later — scoped to one user, customer, or agent, so nothing leaks between them. Underneath sits a real database, so the same store handles source-grounded retrieval too. Here are concrete things people build with it.

An assistant that remembers users across sessions

The problem: Your agent forgets the user between sessions. The chat history is gone, and people repeat themselves every conversation.

What Aether does: Create one Memory per user. Remember preferences and facts as they come up; next session — hours or months later — recall the most relevant ones by meaning and fold them into the prompt.

import os

from aether import Memory

memory = Memory("user-42", api_key=os.environ["AETHER_API_KEY"])
memory.remember("Prefers concise weekly summaries")

# A new session, any time later
for item in memory.recall("communication style", k=3):
    print(item.score, item.text)

The recall runs in a new process on a new day — memory survives the session.

Who builds this: Anyone shipping a chatbot, copilot, or agent that talks to the same person twice.

Start here: Python quickstart or TypeScript quickstart — first memory stored and recalled in five minutes. Same path in Go and .NET.

Managed memory for a fleet of agents

The problem: You run many agents — or one agent product serving many customers — and each needs its own durable context, without copying prompts, vectors, and documents between tools.

What Aether does: Scope a Memory to any stable id — user, customer, patient, agent. Every operation is isolated to that entity server-side; one customer's memory never surfaces in another's recall. Works with Claude, GPT, Gemini, or any LLM provider.

Who builds this: Agent platforms, multi-tenant SaaS copilots, agent fleets.

Start here: Multi-tenant patterns — a partition per end-client, with isolation the server enforces.

Support agents with context across sessions

The problem: Customers re-explain their setup, plan, and open issues every time they contact support, because the bot only knows the current chat.

What Aether does: Remember customer preferences, product facts, and open issues as conversations happen; recall them when the customer returns, so the agent picks up where it left off instead of replaying chat history. Recency-weighted ranking keeps yesterday's report ahead of a stale mention, and fact extraction distills long transcripts into clean, recallable facts.

Who builds this: Support teams, e-commerce companies, B2B SaaS.

Start here: Modeling conversations — threads, entity scoping, and the per-prompt context mix.

Also works for: source-backed retrieval

Memory leads, but the database underneath is a full document store with semantic search. When the job is answering from source material rather than remembering people, the same API covers it:

  • AI chatbot for your docs. Feed in help docs or a wiki; Aether finds the 3-5 most relevant sections and hands them to an LLM, which answers from your actual docs, not its training data. Swap in course material for a grounded AI tutor. Start with the zero-to-chatbot guide — about 30 lines of Python.
  • Research assistant that cites sources. Every result carries a doc_id and a ranking score, so the LLM can cite the document behind each claim. The RAG quickstart shows the retrieve-and-cite pattern.
  • Personal knowledge base. Dump in notes, PDFs, articles, and transcripts; find them later by meaning, not keywords. The CLI reference loads files from the command line.

The strongest builds combine both: a support bot that remembers the customer and retrieves the right knowledge-base article for the same reply — pick your LLM in the integrations overview.

Not sure? Start by remembering one thing

The Python or TypeScript quickstart has an agent remembering and recalling in about five minutes; if document Q&A is your job today, the zero-to-chatbot guide is still the 30-line classic. Once you watch your agent remember something across a restart, you'll know what to build next.