AtomicMemory components for Langflow

Four Langflow custom components backed by the Python atomicmemory SDK:

They appear in the Langflow component sidebar under the atomicmemory category:

Requirements & compatibility

• Python ≥ 3.10, atomicmemory >= 1.0.1, langchain-core.
• Langflow is the host and must be installed in the same environment. Tested with Langflow >=1.6,<2.0 (the components import a few lfx internals; see the loader smoke test). Newer Langflow majors may move these symbols.
• A running AtomicMemory Core (default http://localhost:17350). Core needs an LLM/embeddings key for ingest extraction.

Heads up: ingest runs synchronous LLM extraction + embedding, so storing a memory can take seconds (sometimes ~20s). Writes are explicit (Store Message) so this latency is visible, not hidden. Chat Memory is read-only — it never auto-writes on every turn. If the backend is unreachable, Chat Memory fails closed (raises a clear error) by default; set its Fail open on error toggle to return empty history instead.

Install

pip install atomicmemory-langflow            # into Langflow's environment
# copy the component entry files into your Langflow components root:
npx @atomicmemory/langflow-plugin --target ~/.langflow/components --python <langflow-python>
# or set the components root via env instead of --target:
LANGFLOW_COMPONENTS_PATH=~/.langflow/components npx @atomicmemory/langflow-plugin --python <langflow-python>

Restart Langflow; the components appear under the atomicmemory category.

Scope, identity & multi-tenant safety

Memory is scoped by user (required) and optional session (thread). User ID defaults to the Langflow run user when blank; an explicit value overrides it. Note this is run context, not strong auth — in CLI/anonymous paths Langflow may auto-generate an opaque user id.

Search Context recalls user-scoped (across sessions) by default — long-term memory should persist beyond a single conversation, and Core hard-filters search/list by session. Set its advanced Scope to session toggle to restrict retrieval to the current session. Chat Memory (this-conversation history) and Store Message remain session-aware.

(namespace is not exposed in Phase 1: the AtomicMemory Python provider only applies it on search/package, not ingest/list/delete, so exposing it would silently break store/delete scoping. It returns once the SDK honors it end-to-end.)

Trust boundary: scope is the only memory boundary, and Langflow lets user_id/ session_id be set via flow inputs/tweaks. In shared / multi-tenant / Cloud deployments, control who can edit and run flows — a flow author who sets user_id can read/write that user's memories.

Security

• Put API keys only in the API Key (secret) field — never in Provider Config (it is stored in plaintext in the flow). Provider Config is allowlist-only: only known tuning keys (timeoutSeconds, apiVersion) are accepted; everything else — URLs, keys, and any secret-shaped key (accessToken, clientSecret, …) — is rejected.
• provider is validated: Phase 1 accepts only atomicmemory, even via API/tweaks (the UI dropdown is not the only guard).
• API URL is fail-closed for remote hosts. It must be http(s) and resolve to a local host by default; pointing memory at a non-local endpoint requires the operator (not the flow author) to opt in via ATOMICMEMORY_LANGFLOW_ALLOW_REMOTE=1 or ATOMICMEMORY_LANGFLOW_ALLOWED_HOSTS=host1,host2. This is not full SSRF protection: it does not sandbox the loopback interface, so a flow author can still reach services bound to the Langflow host's localhost/127.0.0.1 (any port). Treat flow authors as trusted, or add network-egress controls, on shared/multi-tenant/cloud deployments.
• Retrieved memory is emitted as ordinary context, never as a system message.

Provider neutrality

provider defaults to atomicmemory (the only Phase 1 tested provider). The architecture is provider-neutral — provider name + provider_config flow to the SDK — but other providers are not yet listed in the dropdown.

Testing & known follow-ups

Unit tests run without a live backend (cd plugins/langflow && python -m unittest discover -s tests); the SDK-contract and Langflow-loader tests exercise the real atomicmemory SDK models and lfx template builder when those packages are installed.

Follow-ups (tracked, not yet in this PR):

• End-to-end lane against a real AtomicMemory Core (Docker + Core + an LLM key): Store Message → Search Context → Delete with synthetic data, with the package installed into a Langflow-compatible venv. Unit tests use fakes/model coercion; this lane would catch integration drift the fakes can't.
• Namespace scoping once the Python SDK honors it on ingest/list/delete (today only search/package), at which point the namespace input returns.
• Branded AtomicMemory icon (vendor logo, like the model providers') — deferred. Each component currently uses a distinct Lucide icon (save / search / messages-square / trash). A real brand mark is a Langflow vendor icon, which per Langflow's docs requires frontend changes (an @/icons/AtomicMemory SVG + forwardRef wrapper + a lazyIconImports entry) and so cannot ship from a Python component bundle — it needs an upstream Langflow PR. Logo SVGs exist under supermem-internal-web/static/img/.