agentbay 1.8.1

Persistent memory for AI agents. Store, recall, and share knowledge across sessions.

AgentBay — Persistent memory for AI coding agents

The memory OS for coding agents. Persistent memory, collaboration, and governance for Claude Code, Codex, Cursor, OpenClaw, and any MCP client.

pip install agentbay

Works locally with zero config. No signup, no API key, no credit card. When you want shared memory across machines or teammates, sign in and the same brain syncs to the cloud.

Quick start

from agentbay import AgentBay

brain = AgentBay()                              # local, zero config
brain.store("User prefers dark mode")           # remember a fact
results = brain.recall("preferences")           # search by meaning

That's it. Memory persists across sessions. Search is hybrid (alias + full-text + vector + RRF fusion) so you don't have to think about how the agent will phrase the recall.

Sync to the cloud — one command

When you want the same brain on your other machines, across agents, or shared with a teammate, run:

agentbay login

This opens your browser, you sign up (free, 30 seconds, no card), and the CLI pushes every local memory to your cloud account in one pass. Future stores and recalls flow to the cloud automatically. Your local SQLite stays intact as an offline cache.

What you unlock:

• Same brain everywhere — same memories on your laptop, server, CI
• Team memory — invite teammates, share project knowledge
• Vector search at scale — 1024-dim embeddings via Voyage AI
• Multi-agent handoffs — Claude Code, Codex, Cursor all reading and writing the same brain

Other useful CLI commands:

agentbay status   # show local memory count + cloud connection state
agentbay sync     # push any new local memories to cloud (after login)

Equivalent in Python:

brain = AgentBay()             # local mode
brain.store("…")               # accumulate locally
brain = brain.login()          # browser opens, migrates, returns cloud brain

Why this exists

Coding agents forget everything between sessions. The architecture decisions, the bugs you debugged together, the conventions in your codebase — gone the moment the context window closes. AgentBay gives your agent a brain that compounds across sessions.

• Local-first install. pip install agentbay and you're done. No account creation or SaaS lock-in. Local mode runs from SQLite after setup; the first vector recall may download the FastEmbed model from Hugging Face unless it is already cached on the machine.
• MCP-native. Drops into Claude Code, Cursor, Codex, OpenClaw, and any other MCP client with one line of config. See the MCP server package.
• Clean upgrade path. Local for solo. Cloud for sync across machines. Teams for collaboration. Projects for multi-agent handoff. Governance for the enterprise stack. Each tier adds value rather than gating the previous one.

Wrap your LLM

If your agent uses an OpenAI-compatible chat completion, AgentBay can wrap it so memory is recalled and stored automatically:

from agentbay import AgentBay
from openai import OpenAI

brain = AgentBay()
client = OpenAI()

response = brain.chat(
    client,
    model="gpt-4o",
    messages=[
        {"role": "user", "content": "what did we decide about the auth flow?"}
    ],
)

The wrapper recalls relevant memories, injects them into the prompt, and stores anything the assistant learned. No manual store/recall needed.

MCP integration (Claude Code, Cursor, Codex)

Add this to your MCP client config:

{
  "mcpServers": {
    "agentbay": {
      "command": "npx",
      "args": ["-y", "aiagentsbay-mcp"]
    }
  }
}

That's the entire setup. Your agent now has access to 76 memory and collaboration tools.

Sign in for cloud sync (optional)

agentbay login

The same brain that lived in ~/.agentbay/ now syncs to the cloud under your account. Memories follow you across machines. Teams and projects unlock for collaboration. Sign-up takes 60 seconds at aiagentsbay.com.

Documentation + links

• Full docs — https://www.aiagentsbay.com/docs
• MCP server (npm) — https://www.npmjs.com/package/aiagentsbay-mcp
• Landing page — https://www.aiagentsbay.com
• Issue tracker — https://github.com/thomasjumper/agentbay-python/issues
• PyPI — https://pypi.org/project/agentbay/

Memory model

AgentBay uses four memory tiers with confidence decay:

Tier	TTL	Use case
Working	24h	Session-local notes, in-flight task context
Episodic	30d	Recent task and conversation context
Semantic	90d	Reusable patterns, architecture facts
Procedural	365d	Long-lived how-to memory

Memories cap at 10K per project on the free tier. Confidence scores decay per tier so older memories down-rank without being deleted — your agent's recall stays sharp without you pruning by hand.

License

MIT. See LICENSE.