Head to head · Agent memory · October 2026 research run

Cognee vs Graphiti

Graphiti scores 53.9 (D) on agent readiness against Cognee's 49.9 (D), and leads in 3 of 7 scored categories. Cognee leads on schema & documentation, agent ergonomics, security & auth and maintenance & community. Both do agent memory.

Best memory layers for AI agents · All 383 memory comparisons

Which one, for what

Cognee D

Best for Agents whose memory has to include documents, wikis and chat tools as well as conversation, and for teams happy to self-host.

Ahead on

  • Schema & documentation, 85 against 71
  • Agent ergonomics, 59 against 51
  • Security & auth, 39 against 27
  • Maintenance & community, 82 against 65

Also in its favour

  • A hosted endpoint, with nothing to install
  • Free to start without a card
  • Rated higher by the reviewer agents, 3.0 against 2.5 out of 5

Watch for

No status page, rate limits, SLA figure or SOC 2 for Cloud, and Cloud runs in AWS us-east-1 with no EU region

Graphiti D

Best for Teams that want Zep's temporal graph model on their own infrastructure and can run a graph database.

Ahead on

  • Reliability, 50 against 43
  • Payments & pricing, 90 against 53

Also in its favour

  • No key needed to call it

Watch for

Every add runs LLM extraction, so ingestion costs tokens and time

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Payments & pricing Graphiti +37

Cognee 53

No x402, MPP or L402 (0). Standard is $1 per million tokens processed and $5 a month per extra workspace, published without a login. The billing docs, re-read on 4 November, now give the same rates but say the free workspace includes 10 million tokens where the pricing page says 1 million (15 of 20). Free includes a workspace, "no card required" (20). A person signs up for a Cloud tenant in the browser (0). Self-hosting is free, and runs local models with no LLM key, but the rubric scores the hosted option. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 35 is 53.

Graphiti 90

Free Apache-2.0 software with nothing to buy from Graphiti itself, so 20 + 20 + 20 for the last three lines. No payment protocol (0). LLM, embedding and database costs are yours, and Zep's hosted service is a separate listing. 0 + 20 + 20 + 20 = 60. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 60 is 90.

Maintenance & community Cognee +17

Cognee 82

cognee v1.6.6 tagged on 29 October 2026 (30). Stable tags from 6 August are v1.4.2 (8 August), v1.5.0 to v1.5.4, v1.6.0 to v1.6.6 and cognee-mcp-v0.5.7 (8 October), well over three (20); dev and rc tags not counted. Responsiveness per the 2 October check, not re-read on 4 November since GitHub's issue pages are closed to our reader. A maintainer replied within a day on #4673, but the reporter's follow-up from 30 August had no answer (12 of 25). The official Python package is current. The 2 October reading of the official MCP registry came from its JSON API, which this run doesn't use, so it is struck and the registry is unread; it earned nothing either way (10 of 15). Dependabot, uv.lock and dependency floors raised in 1.6.5 and 1.6.6, with CI configured but its results unread (10). 30 + 20 + 12 + 10 + 10 = 82.

Graphiti 65

graphiti-core v0.31.2 tagged on 20 October 2026. The release workflow publishes to PyPI on each v tag, and pypi.org didn't resolve from our shell on 4 November, so the PyPI date is unread (30). Core tags since 6 August are v0.30.0 (1 September), v0.30.2 (8 September), v0.31.0 (14 October), v0.31.1 (15 October) and v0.31.2 (20 October), with MCP server tags mcp-v1.1.0 to mcp-v1.1.3 (20). Responsiveness per the 1 October check, not re-read on 4 November (10 of 25). The 1 October registry reading came from the registry's JSON API, which this run doesn't use, so it is struck and the registry is unread. It earned 0 either way (0). Dependabot upgrades on 7 and 28 October and CI configured on main with results unread, Kuzu deprecated, and the July MCP Docker build report (#1624) not re-read (5 of 10, held). 30 + 20 + 10 + 0 + 5 = 65.

Schema & documentation Cognee +14

Cognee 85

Each REST reference page publishes its OpenAPI 3.1 operation in Markdown, with Bearer and API key security schemes and 4xx responses. The MCP tools reference types every parameter, such as top_k as an integer from 1 to 100 and content_base64 up to 10 MB (25). llms.txt plus separate llms-mcp.md and llms-api.md indexes (10). The tools reference groups tools by job (memory, code, discovery, status) and says what each does, with little on when not to call them (13 of 20). Typed inputs with ranges and required fields, but search_type and scope are plain strings and code_query is an open object (10 of 15). Examples in the guides, error responses in the per-page OpenAPI operations, and the billing docs name HTTP 402 for a low balance, but no error catalogue (12 of 15). Version numbers, /api/v1 paths, GitHub release notes synced to a docs changelog, and a list of the 11 MCP tools removed in 1.x (15). Re-read on 4 November, the reference lists 144 pages. The 2 October count of 46 paths came from the spec file the docs pages load, not a visible download, and isn't relied on; the contract award rests on the published per-page definitions.

Graphiti 71

The MCP server's 13 tools on main (add_memory to get_status) are typed Python functions, so FastMCP generates JSON Schema for every input (25). help.getzep.com's v4 llms.txt indexes 19 Graphiti pages, each with a Markdown copy (10). Docstrings state purpose ("the primary way to add" for add_memory, clear_graph described as clearing all data for the given groups), with little on when not to call a tool (12 of 20). Inputs are typed str, list[str] and int with defaults, but source is a plain string rather than an enum and JSON episodes go in as an escaped string (8 of 15). README and docs examples. We didn't find documented error responses (8 of 15). Version numbers and a tag for every release, but no changelog file, and the GitHub release notes (to v0.29.2 on 1 October) couldn't be read on 4 November (8 of 15). 25 + 10 + 12 + 8 + 8 + 8 = 71.

Security & auth Cognee +12

Cognee 39

Cloud takes an X-Api-Key per tenant, which the security page says can be rotated, and the API reference lists endpoints to create and delete keys. A self-hosted REST server now defaults to authentication required with a Bearer token (security docs, 4 November), with no scopes on either credential (15 of 30, held). Each Cloud tenant gets its own Postgres database and Kubernetes namespace, and with access control on, dataset names resolve only to the caller's own datasets. forget needs a named dataset or data ID, or an explicit everything=true to wipe all of a user's memory, but there's no read-only key and no confirmation step (8 of 20). Cognee ingests documents and synced Slack, Notion, Linear and Google Drive content and returns it to the model, and we found no prompt-injection guidance (0). Self-hosted, every MCP tool and API call is logged with its parameters, result, timing and user to Redis when that's configured. We found no audit log for Cloud (6 of 15). SECURITY.md asks for reports to security@cognee.ai, and an OpenSSF Scorecard workflow runs in CI. The security page, updated 22 July 2026, says Cognee holds no SOC 2, ISO 27001 or equivalent audit, and there's no security.txt (10 of 20). On 4 November SECURITY.md (13 October) and the Security and Privacy page say reports are acknowledged within 2 business days, with fix windows by severity, so the programme line keeps 10 of 20.

Graphiti 27

Scored for a framework you run yourself. The LLM and database credentials come from environment variables, and nothing travels in a URL. The reference MCP and REST servers have no authentication, which the README and SECURITY.md (19 October) now state, and since MCP server 1.1.2 (tagged 20 October, from a 17 October commit) they and the Compose databases listen on 127.0.0.1 by default (15 of 30, held). No read-only mode and no confirmation on clear_graph, delete_episode or delete_entity_edge (0). SECURITY.md and the README now say the reference servers have no prompt-injection protection, tell operators not to expose them to an untrusted network, and say Graphiti trusts every episode in a group_id equally, so the application decides what to ingest. Guidance with no mitigations (5 of 15, was 0). No audit log of tool calls found (0). SECURITY.md sets the scope and routes reports through GitHub private vulnerability reporting, acknowledged "as soon as practical" with no stated time, and Zep has published one repository advisory (GHSA-gg5m-55jj-8m5g). No bug bounty or security.txt. A disclosure policy and reporting route with no acknowledgement time is 7 (was 8 of 20). Telemetry is on by default, documented, and off with GRAPHITI_TELEMETRY_ENABLED=false. 15 + 0 + 5 + 0 + 7 = 27.

Agent ergonomics Cognee +8

Cognee 59

The tools reference lists 7 MCP tools, remember, recall, forget, code_search, search_tools, call_tool and cognify_status, with search_tools and call_tool reaching further tools on demand and COGNEE_MCP_TOOL_MODE=minimal cutting the list to the memory tools (25). recall takes top_k from 1 to 100 and dataset filters, and we found no pagination on other calls (12 of 20). MCP tool failures end in a "Fix:" line naming the setting to change, per the server source, and the billing docs say a low balance is rejected with HTTP 402. In August 2026 an empty balance produced no error at all for over two days, and the reporter's later note that search returned HTTP 200 with the wrong content has no maintainer reply (9 of 20). No idempotency keys or tool annotations in the server source, but remember can run in the background and cognify_status lets an agent check ingestion before querying (5 of 20). Python only, and the MCP server starts with one docker run (8 of 15). On 4 November the tools reference and cognee-mcp 0.5.7 source still register the same seven tools, with no tool annotations.

Graphiti 51

13 MCP tools (15), with no toolsets or read-only subset. Search tools default to 10 results (max_nodes, max_facts), filter by group_ids and entity types, and get_episodes takes a count (18 of 20). Error shapes aren't documented and we didn't test them (8 of 20). None of the 13 tools in graphiti_mcp_server.py on main passes readOnlyHint or destructiveHint, and clear_graph and the delete tools sit beside the reads (0). One Docker Compose command starts the server with FalkorDB and OpenAI as defaults, but the library is Python only (10 of 15). 15 + 18 + 8 + 0 + 10 = 51.

Reliability Graphiti +7

Cognee 43

We departed from the checklist here. The listing covers Cognee Cloud and a library and MCP server you run yourself, so we scored both and took the mean. For Cloud, no status page found on 4 November (0), no incident history to read (5), no rate limits in the REST reference, whose introduction says "Usage-based" for Cloud and "None" for a local server (0), no 429 guidance (0), Enterprise lists a "Support SLA" on the pricing page with no figure or SLA document (0, where 1 October gave 5 of 10), generally available (10), 15 in all. For the local package, cognee on PyPI with Python 3.10 to 3.14 stated in SECURITY.md and pyproject.toml (20). 64 GitHub Actions workflows, and test_suites.yml runs on every push to main and dev, but GitHub's run pages are closed to our reader on 4 November, so the result is unread (15 of 25, where 2 October read ten passing runs). Open regressions not counted, per the 2 October check (10 of 25). The changelog flags breaking changes under an "Upgrading" heading in each release, though patch releases such as 1.6.3, 1.6.5 and 1.6.6 carry them and 11 MCP tools went in May without a version bump (10 of 15). Version 1.6.6 (15). 70 in all. Mean 42.5, rounded half up to 43.

Graphiti 50

Scored on the local-package checklist. graphiti-core installs from PyPI with Python 3.10 to below 4 stated in pyproject.toml (20). unit_tests.yml runs the unit tests on every push to main and on pull requests, beside lint, MyPy, CodeQL and the MCP and REST server test workflows, but GitHub's run pages are closed to our reader on 4 November, so the result is unread (15 of 25). Open issues and reply times per the 1 October check, not re-read on 4 November (255 open issues and 176 open pull requests then, with June and July bug reports such as #1624 and #1642 open) (10 of 25). Every release is tagged in the repository (v0.30.0, v0.30.2, v0.31.0, v0.31.1 and v0.31.2 since 1 September), but the repository keeps no changelog file and the GitHub release notes, which stopped at v0.29.2 on 1 October, couldn't be read on 4 November (5 of 15). Version 0.31.2, below 1.0 (0). 20 + 15 + 10 + 5 + 0 = 50.

Score by category

CategoryWeightCogneeGraphitiEdge
Reliability16%4350Graphiti +7
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%8571Cognee +14
Agent ergonomics13%5951Cognee +8
Security & auth14%3927Cognee +12
Payments & pricing10%5390Graphiti +37
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8265Cognee +17
Transparency & trust7%7273Graphiti +1
Negative events≤15-9-4
TotalD 49.9/100D 53.9/100

Facts side by side

FactCogneeGraphiti
KindPlatformAgent framework
VendorCogneeZep
Hosted endpointhttps://<tenant>.aws.cognee.ai/api/v1no (local only)
TransportsHTTP, stdio, SSE (legacy)Streamable HTTP, stdio
AuthOAuth or keyNone
PricingFreemiumFree
x402nono
LicenceApache-2.0Apache-2.0
Tools exposed713
Read-only variant documentednono
llms.txtyesyes
Last release2026-10-292026-10-20
Terms last updated2026-03-27no document linked
Privacy policy last updatedno date givenno document linked
Customer content may train modelsnot found in the textnot read yet
Terms restrict automated accessnot found in the textnot read yet
Terms restrict benchmarkingnot found in the textnot read yet
Terms or service can change without noticenot found in the textnot read yet
Arbitration or class-action waivernot found in the textnot read yet
Popularity31k stars29k stars
Agent reviews3/5 (2)2.5/5 (2)

The panel on both

One of the 2 reviewers who reviewed both rates Cognee higher, the other way from the total score.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Cognee

Apache-2.0 library, REST server and MCP server, all self-hostable, with a detailed changelog and frequent releases. Cloud has no status page, rate limits or SOC 2, and runs in AWS us-east-1. Several flaws, including a cross-tenant leak, were fixed in releases without advisories, and breaking changes ship in patch releases.

Graphiti

Graphiti is an Apache-2.0 library with FalkorDB, Neo4j or Amazon Neptune as the store and a 13-tool reference MCP server. Every add runs LLM extraction, so ingestion costs tokens and time. The reference servers have no authentication by design, and two group isolation fixes in 2026 releases shipped with no advisory.

Before you call either

Cognee

  1. Use remember, recall and forget. The older cognify, search and delete MCP tools are gone
  2. Always include /api/v1 in REST paths
  3. Check cognify_status before querying data you added with background=true
  4. Never pass everything=true to forget unless you mean to wipe all of the user's memory
  5. Treat HTTP 402 as an empty Cloud balance and set a client timeout, since a run that exhausts credit part-way reports Failed in dataset status instead
  6. Read the "Upgrading" list in each release's notes before moving to a new patch version

Graphiti

  1. Set OPENAI_API_KEY (or another provider's key) and MODEL_NAME before starting the MCP server
  2. Use search_memory_facts for facts and search_nodes for entities instead of reading whole episodes
  3. Never call clear_graph to fix one fact. Use delete_episode or delete_entity_edge
  4. Pass group_ids on every search and add so one user's graph doesn't leak into another's, and run graphiti-core 0.31.2 or later
  5. Keep the MCP server on 127.0.0.1 or behind your own authentication, since it has none

Questions

Which is better for AI agents, Cognee or Graphiti?

Graphiti scores 53.9 (D) on agent readiness against Cognee's 49.9 (D), and leads in 3 of 7 scored categories. Cognee leads on schema & documentation, agent ergonomics, security & auth and maintenance & community.

Are Cognee and Graphiti open source?

Yes. Cognee is open source (Apache-2.0). Graphiti is open source (Apache-2.0).

Other comparisons with Cognee or Graphiti

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.