https://raw.githubusercontent.com/OrcVole/cognee-cloudron/main/CloudronVersions.json
Description
1.6.8
Cognee
Cognee is an open-source memory layer for AI agents. It turns documents, conversations and other data
into a knowledge graph together with a vector index, so an agent can recall facts and how they relate.
Add data through the web interface or the REST API, build the graph with cognify, and query it by
chunk search, by graph search, or by asking a question and getting an answer drawn from the graph.
This package runs the API and the web interface behind one address. PostgreSQL comes from the platform; the graph store (Kuzu) and the vector store (LanceDB) are files in the app's data directory, so they are covered by the platform's backups.
A language model is optional. With none configured, Cognee extracts entities with a small local model
(about 784 MB) and embeds text with another (about 67 MB). Both download the first time they are used,
not at install, and are kept in a persistent directory that is left out of backups. Chunk search works this way. Answer-style searches
and richer graph extraction need a language model: point Cognee at any OpenAI-compatible endpoint with
the LLM_* and EMBEDDING_* settings described in the post-install notes.
Sign-in is by email and password only. There is no single sign-on, no mail is sent, and there is no password reset by email: an administrator with access to the app resets passwords. Self-registration is closed, and telemetry is off.
The web interface is upstream's open-source one. Its Overview panels are marked as Cognee Cloud features and stay empty; the Brain, Search and Mindmap pages are the useful ones.
Recent Changes
Update Cognee 1.6.6 to 1.6.8 (commit 069f3cc, includes 1.6.7). Security items in upstream's notes (1.6.7): raw database connection strings and connection strings carrying Cognee's own credentials are refused by default; server file paths sent over HTTP are refused by default;
ACCEPT_LOCAL_FILE_PATHnow also gatess3://inputs; blocked file ingestion answers 403 instead of 415.The graph database engine (Ladybug) moves from 0.19.0 to 0.22.3, and this is a one-way change. Upstream states that 0.22.3 opens stores written by 0.17.1 to 0.19.0 in place; the first start after updating upgrades the graph storage format (code 43 to 49). A test with the previous image confirmed it: Ladybug 0.19.0 refuses the upgraded store ("Database file version: 49, Current build storage version: 43"), so a rollback to package 0.1.2 cannot open it. Take a Cloudron backup before updating. Upstream makes the move because 0.19.0 could exhaust its database size budget during a long cognify and leave the store unopenable (upstream issue 6679). The baked Ladybug extension image is updated to match (it now carries the v0.22.0 JSON extension that 0.22.3 requests). Ladybug's default maximum database size also rises from 32 GiB to 128 GiB; this is address-space accounting, not disk use.
Upstream behaviour changes: on the API server,
ALLOW_HTTP_REQUESTS(ingesting from a web URL) andALLOW_DATABASE_INGESTIONare now off unless set, so URL and database-connection ingestion through the API stop working until an operator sets them in the app's environment. This package keepsACCEPT_LOCAL_FILE_PATH=true, which it already set. Theinstructorlibrary is no longer installed with the core package (the default structured-output path does not use it). The LanceDB (0.39) pin is unchanged.New upstream features, none switched on by this package: a Slack connector and app, integration sync schedules, and a RabbitMQ embedding worker. Two additive database migrations (integration sync policies and a credentials column) run on the first start. Dependency updates: fastembed 0.9.1 and pydantic 2.13 floors, onnxruntime floor lowered, enola-cli 0.5.1, docling 2.116.

