Cognee

View as Markdown

Adds persistent semantic memory to agents through three tools: cognee.remember, cognee.recall, and cognee.memory_status.

This is a tool node (classType: ["tool"], invoke capability, no data lanes). Multiple agents can connect to the same Cognee node and share its operator-configured dataset.

What it does

Cognee turns plain text into a semantic knowledge graph and embeddings. Recall always requests references from Cognee, but this node does not require every returned result to contain references. The graph represents semantic relationships. It is not an AST, import graph, or call graph, and the node does not accept repository URLs for generic source ingestion.

Memory is grouped by dataset. By default, every tool call is locked to the dataset configured on the node. Per-call dataset selection works only when allow_dataset_override is set to true.

Tools

• cognee.remember sends plain text to POST /api/v1/remember, storing and processing it in one operation. Set run_in_background to return while processing continues.
• cognee.recall sends a natural-language query to POST /api/v1/recall. It always requests references and defaults to GRAPH_COMPLETION_DECOMPOSITION.
• cognee.memory_status resolves the selected dataset name to its UUID and checks the Cognify pipeline. It returns pending, running, completed, or failed.

There is no destructive clear/reset tool.

REST API contract provenance

This node follows Cognee's Recall API reference and Get Datasets API reference, checked 2026-07-21. POST /api/v1/recall uses the documented camelCase JSON fields searchType, topK, and includeReferences; this node always sends includeReferences: true. GET /api/v1/datasets is called without pagination parameters and expects Cognee's documented bare JSON list response.

Setup

Configure these node settings:

• base_url: Cognee server URL. Self-hosted Cognee defaults to http://localhost:8000; a containerized engine needs an address reachable from the container. For Cognee Cloud, use the instance URL from platform.cognee.ai.
• api_key: optional for self-hosted servers with access control disabled. Otherwise set the secure field or the COGNEE_API_KEY environment variable. Never commit a key.
• dataset: shared default dataset for all connected agents. Defaults to main.
• allow_dataset_override: permits a tool call to select a different dataset. Defaults to false.
• search_type: default recall strategy. Defaults to GRAPH_COMPLETION_DECOMPOSITION.
• top_k: default recall result limit from 1 through 100. Defaults to 15.
• request_timeout: per-request timeout from 5 through 600 seconds. Defaults to 120.

Shared-agent example

See the copyable cognee-shared-memory-agents.pipe pipeline. A RocketRide component can have multiple tool/control entries, so one tool_cognee node can be controlled by both the writer and researcher agents. Prefer one shared Cognee node to duplicating it for each agent: one component keeps the server, credentials, dataset, and retrieval settings from drifting apart.

Shared recall requires all participants to use the same Cognee server, an authenticated identity with permissions to the shared memory, and the same dataset. Leave allow_dataset_override disabled to keep every call in the operator-controlled dataset; per-call dataset drift is rejected unless the setting is explicitly enabled.

Wiring agents to Cognee does not automatically copy their prompts, transcripts, or tool results. An agent must explicitly call cognee.remember to create durable memory, and another agent must explicitly call cognee.recall to retrieve it. Each agent_rocketride still needs its own memory_internal control for run-scoped working memory. Cognee adds durable cross-agent memory; it does not replace that working memory.

Simultaneous writes from shared agents may overlap. This node provides no transaction, serialization, or ordering guarantee across tool calls, and recall is eventually consistent with processing. Prefer a sequential handoff: let the storing agent finish, wait for cognee.memory_status to report completed, and only then have the receiving agent recall.

To test the example sequentially:

1. Set COGNEE_BASE_URL, COGNEE_API_KEY, and ROCKETRIDE_ANTHROPIC_KEY, then open the example.

2. Send this prompt. It gives the writer only a research subject, not the fact that must cross the agent boundary:

   Ask the researcher to identify one lesser-known fact about Grace Hopper and store it in shared
   memory. The researcher must return only a storage confirmation, not the fact. After processing
   completes, recover the fact yourself from shared memory and tell me what it is.

   The researcher independently determines the fact, calls cognee.remember, and returns only a storage confirmation.

3. If remember ran in the background, have the writer call cognee.memory_status until the shared dataset reports completed.

4. Have the writer call cognee.recall for the fact and produce the final response from recalled memory, without receiving the fact directly from the researcher.

Limits

• remember accepts plain text only and is not retried automatically because retrying a write could duplicate ingestion.
• recall requests references and is also single-attempt because it may perform generation.
• Dataset discovery and status checks are idempotent GETs and use RocketRide's shared retry helper.
• memory_status requires the dataset to exist and resolve by exact name.
• Per-call dataset values are rejected unless allow_dataset_override is enabled.
• No destructive dataset operation is exposed to an agent.

Troubleshooting

• Dataset not found: verify the exact configured dataset name and that remember created it.
• Dataset override rejected: enable allow_dataset_override, or omit the per-call dataset so the configured dataset is used.
• Recall is empty: wait for memory_status to report completed, then query the same dataset.
• 401 / 403: set a valid API key for Cognee Cloud or an access-controlled self-hosted server.
• 402: the Cognee token budget is exhausted; add capacity before retrying paid processing.
• Connection refused / timeout: verify base_url from the engine's network environment.

Upstream docs

• Cognee docs: https://docs.cognee.ai
• Cognee API reference: https://docs.cognee.ai/api-reference/introduction
• Cognee Cloud: https://platform.cognee.ai

Schema

Field	Type	Description	Default
cognee.allow_dataset_override	boolean	Allow dataset override Allow agents to select a dataset per call. Leave off when several agents should share one operator-controlled memory scope.	false
cognee.api_key	string	API Key Cognee API key sent as the X-Api-Key header. Falls back to COGNEE_API_KEY. Self-hosted servers with access control disabled can leave this blank.	""
cognee.base_url	string	Server URL Base URL of the Cognee server. Self-hosted Cognee defaults to http://localhost:8000. For Cognee Cloud, use the instance URL from platform.cognee.ai.	"http://localhost:8000"
cognee.dataset	string	Dataset Default Cognee dataset used by all agent tools. Tool calls stay in this operator-controlled scope unless per-call overrides are enabled.	"main"
cognee.request_timeout	integer	Request timeout (s) Maximum seconds to wait for one Cognee request.	120
cognee.search_type	string	Default recall strategy How recall retrieves memory. Graph completion decomposition is the semantic default; the agent can override it per call.	"GRAPH_COMPLETION_DECOMPOSITION"
cognee.top_k	integer	Recall results (top_k) Maximum number of results recall retrieves per query.	15

Dependencies

• requests >=2.34.2
• tenacity