The memory OS for coding agents.
Persistent memory, collaboration, and governance for Claude Code, Codex, Cursor, OpenClaw, and your own agents.
Install AgentBay in this project as the memory OS for coding agents.
Use local SDK install first:
pip install agentbay
npm install agentbay
If MCP is available, connect:
https://www.aiagentsbay.com/api/mcp
Authorization: Bearer $AGENTBAY_API_KEY
Verify by storing one PATTERN memory, then recall it before editing code.Install for your stack
Pick your agent. One command each.
Claude Code
Hosted HTTP MCP for Anthropic CLI.
Cursor
Drop-in MCP config for the Cursor editor.
Codex
Hosted HTTP MCP for OpenAI Codex agents.
OpenClaw
One-shot installer with the OpenClaw plugin.
Windsurf
MCP config for Codeium Windsurf.
Hermes
Protocol-level install for Hermes agents.
from agentbay import AgentBay
brain = AgentBay() # works locally, no signup
# Your agent remembers across sessions
brain.store("User prefers dark mode", user_id="thomas")
brain.recall("preferences", user_id="thomas")
# Wrap your LLM — memory is automatic
response = brain.chat([
{"role": "user", "content": "fix the auth bug"}
])
# Recalls relevant memories → injects into context → stores learningsNever repeat yourself to your agent again.
Every session picks up where the last one left off.
fewer tokens wasted
~400 tokens per recall vs 5,000+ in system prompts
first — runs fully offline
SQLite-backed local mode. No signup, no cloud, no API key.
config to start
pip install, import, done. No API key needed.
See it in action
When you need more than one agent
Teams — intelligent knowledge sharing
Auto-share policies propagate pitfalls across your fleet. Nightly team dreaming finds cross-agent contradictions. 5-role permissions keep things safe.
team.set_policy(auto_share_types=["PITFALL", "DECISION"])
# Now every PITFALL stored by any agent is auto-shared
# Team dreaming finds contradictions overnightProjects — Multi-agent collaboration
Shared knowledge, task dependencies, and structured handoffs between agents. One agent picks up where another left off.
# Agent A stores knowledge + hands off
proj.record("Auth uses JWT with refresh tokens", type="PATTERN")
proj.handoff(summary="Auth module complete", next_steps=["Add rate limiting"])
# Agent B picks up
context = proj.onboard() # gets knowledge, tasks, handoff notesConnect a whole team's agents — no shared API keys.
AgentBay speaks OAuth 2.1 over MCP. Connect Claude Code or Claude Desktop with a scoped, revocable grant instead of pasting a long-lived key into a config file. Production beta.
Start local. Go cloud when ready.
All your memories migrate with one command. Nothing lost.
# Ready for cloud? Teams? Projects?
brain = brain.login()
# Opens browser → sign up → paste key → done
# All local memories migrate automaticallyFree cloud tier: 10,000 memories and 1,000 API calls/day. No credit card required.