← Back to research
·12 min read·company

Cognee

Cognee is an open-source AI memory engine that turns scattered data into a queryable knowledge graph via its ECL (Extract-Cognify-Load) pipeline, unifying relational, vector, and graph storage. $7.5M seed led by Pebblebed in February 2026, 17.7K+ GitHub stars, and 1M+ pipeline runs — 500x year-over-year growth.

Key takeaways

  • $7.5M seed led by Pebblebed (Pamela Vagata, ex-OpenAI; Keith Adams, ex-Facebook AI Research) in February 2026, with 42CAP, Vermilion Ventures, and angels from Google DeepMind, n8n, and Snowplow
  • The differentiator is the ECL (Extract-Cognify-Load) pipeline: data from 38+ sources is structured into a knowledge graph with embeddings, then a "memify" layer applies feedback-driven refinement — the self-improving part — over unified relational, vector, and graph storage
  • Traction inflected hard: pipeline runs grew from about 2,000 in 2025 to over 1 million in 2026 (500x), with 70+ companies deployed including Bayer for scientific research workflows
  • Apache-2.0 open source with 17.7K+ GitHub stars and a hosted cloud from $35/month, so the self-host-to-managed path is cheap to start and auditable end to end

FAQ

What is Cognee?

Cognee is an open-source memory engine for AI agents that ingests scattered data and builds a persistent, self-improving knowledge graph agents can query across sessions.

How much does Cognee cost?

The open-source engine is free under Apache-2.0; the hosted cloud runs $35/month (Developer, 1,000 documents), $200/month (Cloud, 2,500 documents, 10 users), with custom-priced on-prem enterprise deployment.

How does the ECL pipeline work?

Extract pulls data from 38+ sources, Cognify structures it into a knowledge graph with embeddings and relationships, and Load serves it to agents — with a "memify" layer refining the graph from feedback over time.

How is Cognee different from Zep?

Both build graph-based agent memory, but Zep centers on a temporal knowledge graph of conversational facts, while Cognee is a general data-to-memory pipeline that unifies relational, vector, and graph storage and emphasizes bulk document ingestion.

Executive Summary

Cognee is an open-source memory engine for AI agents: it ingests scattered data — documents, databases, conversations — and turns it into a persistent, self-improving knowledge graph that agents query across sessions.[1][2] The mechanism is its ECL pipeline — Extract, Cognify, Load — which pulls from 38+ data sources, structures content into a graph with embeddings and relationships, and layers a "memify" step that refines the graph from feedback over time; underneath, it unifies relational, vector, and graph storage rather than bolting a vector index onto a chat log.[3] The pitch to the agent-self-improvement category is that memory should be a structured, evolving substrate, not a pile of retrieved snippets.

Berlin-based Topoteretes (founded 2024 by CEO Vasilije Markovic; the repo dates to August 2023) closed a $7.5M seed on February 19, 2026, led by Pebblebed — the fund of OpenAI co-founder Pamela Vagata and Facebook AI Research founder Keith Adams — with 42CAP, Vermilion Ventures, and angels from Google DeepMind, n8n, and Snowplow.[3][1] The traction curve behind the raise: pipeline runs grew from roughly 2,000 in 2025 to over 1 million in 2026 — 500x year over year — across 70+ company deployments including Bayer, which uses Cognee for scientific research workflows.[3] The repo stands at 17.7K+ stars as of June 2026, up from the 12K+ cited at the raise.[1][3]

AttributeValue
CompanyTopoteretes UG (Berlin)[3]
FounderVasilije Markovic (CEO); early co-creators Boris and Laszlo per the Show HN posts[3][4]
Founded2024 (repository created August 2023)[3][1]
Funding$7.5M seed (Feb 19, 2026) led by Pebblebed; 42CAP, Vermilion Ventures, angels[3]
GitHub Stars17.7K+ (June 2026); 1,880+ forks[1]
LicenseApache-2.0; primary language Python[1]
Named UsersBayer, University of Wyoming, Dilbloom, dltHub[3]

Product Overview

The core loop: point Cognee at your data, call cognify, and get back a knowledge graph that agents query for grounded, relationship-aware answers instead of raw chunk retrieval. Documents become both searchable by meaning (embeddings) and connected by typed relationships (graph), and the "memify" layer applies feedback-driven refinement so retrieval quality compounds with use — the self-improvement claim in concrete form.[3][2] Memory separates into a session layer (short-term working memory) and a permanent layer (long-term knowledge artifacts such as user data and interaction traces), continuously cross-connected inside the graph.[2]

The team's own evaluation, published as a paper alongside the June 2025 Show HN launch, reported nearly 90% on standard industry benchmarks, with the eval harness in the open repository.[4][5]

Key Capabilities

CapabilityDescription
ECL pipelineExtract → Cognify → Load: ingestion from 38+ sources into a structured memory graph[3]
Unified storageRelational, vector, and graph databases behind one abstraction[3]
Memify refinementFeedback-driven layer that improves the graph over time[3]
Node setsTag-like grouping for filtering and multi-tenant memory separation, with per-user graphs and permissions[4]
Graph completion retrievalGraph-aware retrieval with weighting and self-improving feedback[6]
Integrated evalsEvaluation tooling ships with the platform, including the free tier[7][5]

Product Surfaces

SurfaceDescriptionAvailability
Python libraryOpen-source engine, self-hostedGA, Apache-2.0[1]
Cognee CloudHosted API on AWS, GCP, and AzureGA, from $35/month[7]
On-premPrivate-cloud or on-premises enterprise deploymentCustom[7]

Technical Architecture

Cognee is a Python engine that orchestrates three storage families — relational, vector, and graph — behind a single pipeline abstraction, so teams can swap underlying databases without rewriting memory logic.[1][3] Development is active: 166 contributors (including anonymous) per the GitHub API, release v1.1.2 shipped May 30, 2026, and commits landed the day of this profile.[1] The seed round's stated roadmap points at the architecture's next moves: a Rust engine for edge and on-device agents, expanded multi-database support, and 30+ new connectors.[3]

Key Technical Details

AspectDetail
DeploymentSelf-hosted (open source), hosted cloud (AWS/GCP/Azure), or on-prem enterprise[1][7]
Model(s)Bring-your-own LLM for extraction and cognify steps; model-agnostic[1]
StoragePluggable relational + vector + graph backends, unified[3]
Open SourceApache-2.0, Python, 17.7K+ stars, 1,880+ forks, 166 contributors[1]

Strengths

  • Verified hockey-stick usage — pipeline runs went from ~2,000 in 2025 to 1M+ in 2026 (500x), a trajectory corroborated by the founder's mid-2025 HN job post citing 116K runs/month and "projecting north of a million."[3][6]
  • Credible deep-tech backing — Pebblebed is run by an OpenAI co-founder (Pamela Vagata) and Facebook AI Research's founder (Keith Adams); angels span Google DeepMind, n8n, and Snowplow.[3]
  • A real enterprise anchor — Bayer runs Cognee in scientific research workflows, with the University of Wyoming, Dilbloom, and dltHub also named, across 70+ company deployments.[3]
  • Open core with a cheap managed on-ramp — Apache-2.0 engine plus a $35/month hosted tier means evaluation costs nothing and migration paths run in both directions.[1][7]
  • Published evals, not just claims — the team shipped a paper (arXiv 2505.24478) and keeps the benchmark harness in the repo, unusual transparency in a category with disputed benchmarks.[5][4]

Cautions

  • LLM-in-the-loop economics — like Mem0 and Zep, Cognee sends text through an LLM for extraction and structuring, which competitors attack on cost at scale; budget for per-document API spend on top of subscription.[6]
  • Temporal evolution is unfinished — when a Show HN commenter asked how Cognee handles facts that change over time, the founder answered that temporal resolution mechanisms "are being built"; as of the June 2025 thread this remained roadmap.[4]
  • Thin independent community signal — the best-performing Show HN drew 9 points and 2 comments; most substantive HN mentions of Cognee come from founders of competing tools or from Cognee's own founder.[6][4]
  • Document-count pricing climbs fast — the hosted tiers meter by documents (1,000 at $35/month), and add-on packs run to $750 for 15,000 documents, which can surprise teams with large corpora.[7]
  • Crowded, contested category — agent memory has many funded rivals (Mem0, Zep, Letta) plus newer entrants, and independent reviewers note most systems converge on extract-store-retrieve with weak mechanisms for deciding what is worth remembering.[6]

What Developers Say

Community discussion is modest as of June 2026 — Cognee's Show HN threads drew single-digit comment counts, and the most detailed independent commentary comes from an eight-system comparison and from builders of competing memory tools (read those as adversarial). Founder Vasilije Markovic (vasa_) is active in HN memory threads, so some framing is vendor voice.[6][4]

"Hey, this looks super interesting - nice work! ... How does cognee handle temporal evolution of memory (e.g. when facts change over time)?" — an HN commenter on the Show HN launch[4]

"every AI memory system I tried (mem0, Cognee, Zep) makes 2-3 LLM API calls just to store a single memory." — a competing memory-tool builder on Hacker News[6]

"Structure without selection pressure is art. Many of these systems build elaborate relationship schemas with no mechanism to decide what's worth remembering." — an HN commenter who prototyped eight memory systems including Cognee[6]


Pricing & Licensing

TierPriceIncludes
Free$0Build and run memory workflows, auto-generated knowledge structures, integrated evals, 28+ data sources, community support
Developer$35/month1,000 documents or 1GB, 1 user, hosted on AWS/GCP/Azure, 10,000 API calls
Cloud$200/month2,500 documents or 2GB, 10 users, multi-tenant architecture, per-user/domain memory grouping, dedicated Slack channel, 10,000 API calls
On-Prem (Enterprise)CustomOn-premises or private cloud, architecture review, premium support/SLA, AI FDE engineers, roadmap prioritization

Add-on document packs: +1,000 for $35, +3,000 for $100, +15,000 for $750. All pricing as of June 2026.[7]

Licensing model: Apache-2.0 open-source engine; hosted cloud and on-prem enterprise are commercial.[1][7]

Hidden costs: Extraction and cognify steps consume your own LLM API tokens on top of subscription fees; document-metered tiers mean large corpora require add-on packs.[6][7]


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
ZepThe closest graph-memory rival — Zep's temporal knowledge graph (Graphiti) specializes in tracking conversational facts as they change over time, exactly the temporal-evolution problem Cognee lists as in-progress; Cognee counters with broader data-to-memory pipelines and unified tri-store architecture
Mem0The category's mindshare leader by GitHub stars; minimal-structure memory layer (LLM extraction, no rich schema) versus Cognee's full knowledge-graph construction — simpler to adopt, less structured to reason over
HindsightRuns four-way parallel retrieval with cross-encoder reranking on a single PostgreSQL database — a one-database simplicity bet against Cognee's pluggable relational + vector + graph stack
LettaA full stateful agent runtime with tiered memory rather than a memory engine you attach to your own agents

When to Choose Cognee Over Alternatives

  • Choose Cognee when: the memory problem is document- and data-heavy — turning corpora, databases, and multiple sources into one queryable graph — and you want an Apache-2.0 core you can self-host with a managed escape hatch.
  • Choose Zep when: conversational memory with first-class temporal fact tracking (what was true when) is the center of the workload.
  • Choose Mem0 when: you want the simplest framework-agnostic memory layer and the largest community, without knowledge-graph overhead.
  • Choose Hindsight when: you want agent memory that lives entirely in PostgreSQL you already operate.

Ideal Customer Profile

Best fit:

  • Teams building agents over large, heterogeneous corpora — research, scientific, legal, enterprise documents — where relationship-aware retrieval beats chunk similarity (the Bayer pattern)[3]
  • Engineering organizations that require a self-hostable, Apache-2.0 memory layer for data-residency or audit reasons[1]
  • Builders who want to customize graph ingestion, generation, and retrieval logic rather than accept a fixed memory schema[6]

Poor fit:

  • Chat products whose memory needs are simple user-preference recall — lighter layers cost less per stored memory
  • Workloads dominated by rapidly changing facts where temporal resolution is the core requirement, until Cognee's in-progress mechanisms ship[4]
  • Token-cost-sensitive deployments storing memories at very high volume, given LLM-in-the-loop extraction[6]

Viability Assessment

FactorAssessment
Financial HealthSolid for stage — $7.5M seed (Feb 2026) led by Pebblebed with notable angels[3]
Market PositionCredible challenger — 17.7K+ stars and a Bayer anchor, but Mem0, Zep, and Letta hold more mindshare in a crowded category[1][3]
Innovation PaceHigh — v1.1.2 in May 2026, daily commits, published evals, and a funded roadmap (Rust edge engine, 30+ connectors)[1][3]
Community/EcosystemGrowing but thin on independent voices — 166 contributors yet single-digit HN engagement on launches[1][6]
Long-term OutlookDepends on converting 500x usage growth into paid cloud/enterprise revenue before better-funded memory rivals consolidate the category[3]

The 500x year-over-year pipeline-run growth is the standout number, and it is partially corroborated by the founder's own contemporaneous HN posts rather than only the funding announcement.[3][6] The structural question is positioning: Cognee's tri-store, graph-first architecture is more ambitious than the lightweight memory layers, but ambition cuts both ways — heavier to adopt, and competing against both simpler tools (Mem0) and deeper specialists (Zep on temporal facts).[6]


Bottom Line

Cognee is the strongest open-source bet on memory-as-knowledge-graph: a genuinely permissive Apache-2.0 engine, published evaluations, an enterprise anchor in Bayer, and a usage curve (2K to 1M+ pipeline runs in a year) that earned a $7.5M seed from ex-OpenAI and ex-FAIR investors. The trades are LLM-in-the-loop costs at scale, temporal fact-tracking that still trails Zep, and a community footprint where most of the loudest commentary comes from competitors.

Recommended for: Teams turning large multi-source corpora into agent memory, especially with self-hosting or data-residency requirements; builders who want to own and customize their memory graph's construction and retrieval logic.

Not recommended for: Simple preference-recall chat memory, high-volume cost-sensitive memory writes, or workloads where temporal fact evolution is the core problem today.

Outlook: Watch whether the Rust edge engine and 30+ promised connectors ship on the funded roadmap, whether temporal resolution lands as a first-class feature, and whether the 1M+ pipeline runs convert into paying cloud customers — the metric the next round will be priced on.


Research by Ry Walker Research • methodology