Manage memory bases

A memory base stores the long-term chat history for a Langflow agent in a vectorized format, enabling semantic retrieval of past conversations across sessions.

Unlike the Message History component, which retrieves messages in chronological order from the messages table, a memory base embeds messages into a vector store and retrieves by semantic similarity to return the most relevant context instead of the most recent messages.

Unlike a knowledge base, which you populate manually with files, a memory base fills itself from a flow's chat sessions, which are stored in Langflow's messages table.

Memory bases and knowledge bases share the same vector-store backends. Configure those backends in Settings → DB Providers. By default, new memory bases use local Chroma. You can also use Chroma Cloud, OpenSearch, or Postgres pgvector, depending on which providers you have configured. For setup steps, see Configure vector database providers.

After a memory base is created, you can include its knowledge in a flow by adding the Memory Base component.

Configure memory bases ingestion

Each memory base can only be attached to one flow, but a single flow can have multiple memory bases attached to it by including multiple Memory Base components. Each time the flow completes a run, the captured messages are counted toward a configurable threshold. When the number of unprocessed runs reaches the threshold, an asynchronous ingestion job writes the messages to the vector store.

Auto-capture controls whether this counting and triggering happens automatically. Auto-capture is enabled by default, so the memory base tracks every flow run and fires an ingestion job when the batch threshold is reached. When auto-capture is disabled, the counter never increments and no ingestion job fires automatically. If auto-capture is re-enabled after being disabled, the memory base catches up to the current state by processing all messages that accumulated while it was disabled.

tip

The flow must contain a Chat Input or a Chat Output component to write messages to the messages table.

By default, the Memory Base component retrieves only chunks that match the current session_id. Disable Filter by Session to search across all sessions that have been ingested into the memory base.

When optional LLM preprocessing is enabled, messages are passed through an LLM before they are embedded. The LLM distills the raw conversation into a compact summary, and the summary is written to the vector store instead of the raw messages. Preprocessing requires an embedding model. For more information, see Preprocessing prompt examples.

Configure a kill phrase to gate ingestion. If the LLM's response contains the configured kill phrase, the batch is skipped, with no data written to the vector store. The ingestion cursor still advances so the same batch is not re-evaluated. The default kill phrase is NO_INGEST: if the LLM response contains NO_INGEST, the batch is skipped. For more information, see Preprocessing prompt examples.

Preprocessing prompt examples

This section includes preprocessing prompts you can copy and paste directly into Langflow. The instructions tell the preprocessing LLM how to filter and summarize each batch of chat history before it is embedded, not how to respond to the user.

To summarize content for support, troubleshooting, or any flow where you want concise facts from each batch, use the general prompt:

Summarize the key facts from this conversation: the issue the user reported,
any troubleshooting steps taken, and the outcome or resolution.
Focus on information a support agent would need in a future session.

To use the kill phrase, instruct the LLM to return it when the conversation contains nothing worth storing. For example:

Summarize the key facts from this conversation: the issue the user reported,
any troubleshooting steps taken, and the outcome or resolution.
Focus on information a support agent would need in a future session.
If the conversation contains no meaningful facts (for example, it is only
greetings or filler), respond with exactly: NO_INGEST

When the LLM returns NO_INGEST, the batch is skipped.

For agent memory that retains durable user context and skips noise, use the context extraction rubric:

Analyze this conversation batch and decide whether it contains durable,
high-value information worth storing in long-term memory.

Award +1 point for each pillar present:
- Core user traits (biography, values, relationships)
- Durable preferences (tools, formats, constraints)
- Active projects and milestones (goals, deadlines, multi-step work)

The batch has zero value if it consists entirely of:
- Transactional tech noise (stack traces, syntax-only debugging)
- Conversational filler ("thanks", "got it", pleasantries)
- Stale re-statements of facts already established

If the total score is 0, respond with exactly: NO_INGEST

If the score is 1 or higher, extract and summarize only the durable
facts in concise bullet points. Omit ephemeral debugging details.

Create a memory base

A memory base is scoped to a single flow and automatically captures that flow's conversation history. Create one from within the visual editor.

1. To open the Memory Bases pane, in the left sidebar, click Memories.

2. Click Create.

3. In the Create Memory pane, enter the following information:

   • Name: Display name for this memory base. Must be unique within your account.
   • Embedding Model: The embedding model used to vectorize messages. The available models are those you have configured in Settings → Model Providers.
   • Vector Database: The DB Provider where this memory base stores vectors. The field defaults to the active provider from Settings → DB Providers. You must select a configured provider before you can create the memory base. Once created, you cannot change the vector database for that memory base.

   Optionally, enable LLM Preprocessing. This option allows you to include a natural language prompt for preprocessing instructions, such as Summarize the key facts from this conversation: the issue the user reported, any troubleshooting steps taken, and the outcome or resolution.

   • Preprocessing Model: The LLM used for distillation.
   • Preprocessing Instructions: Optional system prompt prepended to each preprocessing request. If no prompt is included, the connected LLM processes the conversation however it deems reasonable.

4. Click Create Memory. Langflow creates the memory base and provisions storage on the selected vector database. For local Chroma, that includes an on-disk directory; for remote providers such as Chroma Cloud, OpenSearch, or Postgres pgvector, vectors are stored in the configured cluster or database. A new memory base's status is Empty until the first ingestion job completes.

Use a memory base in a flow

After creating a memory base, add the Memory Base component to your flow to retrieve relevant context from it. For example, to connect a Memory Base to an Agent to give the agent memory of all past conversations, do the following:

1. Create a Simple Agent starter flow.

2. In the visual editor, add a Memory Base component.

3. In the Memory Base field, select the memory base you created in Create a memory base.

4. In the Memory Base component, enable Tool Mode.

5. Connect the Memory Base component's Toolset output to the Agent component's Tools input.

   Each time the flow runs, the component queries the memory base for chunks similar to the current Search Query. By default, only chunks from the same session_id are returned. To retrieve chunks from all sessions, disable Filter by Session.

For more information, see the Memory Base component reference.

Manage memory bases

To view and manage the memory bases attached to the current flow, on the visual editor page, click Memories.

Click a memory base name to open its detailed view. The following information is available for each memory base:

• Name
• Embedding model
• Vector store (the DB Provider selected at memory base creation)
• Ingestion threshold
• Auto-capture
• Session history

To view the embedding model, provider, and vector store for a memory base, open its detail view and expand the configuration summary.

The detail view header contains the following controls:

The session dropdown menu switches between sessions to view their processed message counts and chunk totals.

To reload the session and message data displayed on the page, click Reload sessions and messages.

To toggle Auto-capture, click Auto-capture.

To delete the memory, click Delete.

Memory base storage location

The memory base vectors storage location depends on the DB Provider you selected when you created it. Backend and embedding settings for each memory base are stored on its backing knowledge base record in the Langflow database.

For local Chroma, memory bases use the same on-disk layout as knowledge bases. The default storage location depends on your operating system and installation method:

• Langflow Desktop:
  • macOS: /Users/<username>/.langflow/knowledge_bases
  • Windows: C:\Users\<name>\AppData\Roaming\com.LangflowDesktop\knowledge_bases
• Langflow OSS:
  • macOS/Windows/Linux/WSL with uv pip install: <path_to_venv>/lib/python3.12/site-packages/langflow/knowledge_bases
  • macOS/Windows/Linux/WSL with git clone: <path_to_clone>/src/backend/base/langflow/knowledge_bases

If you set the LANGFLOW_CONFIG_DIR environment variable, the knowledge_bases subdirectory is created relative to that path.

To change the default directory, set LANGFLOW_KNOWLEDGE_BASES_DIR:

export LANGFLOW_KNOWLEDGE_BASES_DIR="/path/to/parent/directory"

For remote providers such as Chroma Cloud, OpenSearch, or Postgres pgvector, vectors are stored in the configured cluster or database, not under the local knowledge_bases directory. Configure those providers in Settings → DB Providers. For setup steps, see Configure vector database providers.

See also

• Memory Base component
• Configure vector database providers
• Manage vector data
• Memory management options