Open-source shared memory infrastructure for multi-agent, multi-device AI workflows.
agent memory hub helps local AI agents share durable memory safely across machines without syncing private runtime state, raw transcripts, credentials, or agent-specific databases.
The lifecycle:
agent learns
-> inbox
-> curator
-> canonical memory
-> cloud handoff
-> downstream refresh
-> agent cold-start perception
python3 -m pip install -e .
agent-memory setup --workspace my-agent-memory --machine laptop --adapter codex
agent-memory register-agent \
--machine laptop \
--agent codex \
--adapter codex \
--primary-memory '~/.codex/memory/shared.md'
agent-memory bootstrap --machine laptop --agent codex
agent-memory sync --machine laptop --agent codex --install-marker
agent-memory trigger "保存到共享记忆"
agent-memory trigger "拉取一下云端的记忆"
agent-memory trigger "记住这个"
agent-memory inbox-add \
--machine laptop \
--agent codex \
--type lesson \
--scope global \
--applicability all_agents \
--fact "Shared memory should be curated." \
--why "Direct canonical edits can conflict." \
--evidence "first setup" \
--destination memory/lessons.md
agent-memory status
agent-memory curate-dry-run
agent-memory curate-apply --machine laptop --agent codex
agent-memory cloud-status
agent-memory cloud-save --message "Update shared memory"
agent-memory cloud-push
agent-memory refresh --machine laptop --agent codex
agent-memory refresh --machine laptop --agent codex --applyThis repository is the open-source framework. User-specific memory state is created by agent-memory setup.
Shared memory preserves source perspective. Inbox notes record observer and
applicability; downstream sync renders the current receiver so an agent does
not mistake another instance's memory for first-person experience.
Prefer facts such as For Codex-family agents... over ambiguous wording such as
Codex's role....