Release Notes - v1.1.0

Release Date: 2026-05-16
Changes: v1.1.0.dev1 → main

Pull Requests: #2848

Summary

Cognee 1.1.0 introduces the Global Context Index, initial Postgres multi-user graph support, and a number of stability, performance, and developer experience improvements. This release also simplifies backend access control configuration, tightens embedding reliability, and adds better cache and concurrency controls to avoid resource exhaustion when running many datasets.

Highlights

• Global Context Index: new memify feature that builds shared context summaries to improve retrieval and summarization.
• Initial Postgres multi-user graph support — groundwork for multi-tenant graph databases.
• Simpler auth posture and safer defaults: single switch for backend access control plus explicit auth override.
• More robust embeddings and subprocess handling (retries, format fixes, watchdog and eviction fixes).
• New dataset concurrency & adapter LRU cache controls to limit resource usage and avoid runaway processes.

Breaking Changes

• Auth posture / access control behaviour changed: ENABLE_BACKEND_ACCESS_CONTROL is now the canonical single switch that determines multi-tenant vs single-user behavior, with REQUIRE_AUTHENTICATION available as an explicit override for the auth requirement. Startup now logs the resolved "auth posture" — check your .env settings and migration steps if you rely on previous defaults.
• Adapter LRU eviction may now shut down subprocess workers when evicted (in subprocess mode). If you run many datasets concurrently, set DATABASE_MAX_LRU_CACHE_SIZE and/or DATASET_QUEUE_MAX_CONCURRENT to appropriate values to avoid unexpected worker restarts.
• Ollama and llama-cpp integrations/dependencies were reworked/removed for some OS/MCP targets. If you rely on these adapters locally or in CI, verify your environment and model adapter availability after upgrading.
• Initial Postgres multi-user support changes DB/connection handling for graph datasets — operators moving from single-tenant setups should review the PostgresGraphDatasetDatabaseHandler and connection configuration.

New Features

• Global Context Index: added end-to-end support for building, assigning, and summarizing global context buckets used during retrieval and memification. This includes new tasks, bucket assignment logic, persistence, and summary-generation components.
• Memory profiles and agent ingestion: memify can now ingest agent and memory profiles to capture richer context during indexing.
• Initial Postgres graph multi-user support: new PostgresGraphDatasetDatabaseHandler and related plumbing to enable multi-tenant Postgres-backed graph datasets.
• CLI improvements: --api-url option now exposes missing API functions and a login flow was added to make authenticated CLI usage easier.
• MCP (micro control plane) UI and server additions: initial web UI/app files and server updates to improve local MCP development and operations.

Improvements

• Cache and concurrency controls: added DATABASE_MAX_LRU_CACHE_SIZE and DATASET_QUEUE_MAX_CONCURRENT settings and clearer .env.template documentation. LRU cache now caps engine instances and (in subprocess mode) the number of concurrent child workers.
• Dataset queue: semaphore-backed queue to limit how many datasets run concurrently by default; prevents resource exhaustion and coordinates with the adapter LRU cache.
• Embeddings reliability: FastEmbed and OpenAI-compatible embedding engines now have better retry behavior, dimension checks at connection time, and fixes for embedding formats — fewer failed ingests and clearer warnings.
• Subprocess stability and diagnostics: improved watchdog handling (won't kill workers when parent is PID 1), buffered-drain and exitcode diagnostics, faulthandler enabled, and safer eviction behavior.
• Database robustness: WAL corruption handling for disk-backed graph DBs (Ladybug), fixes for delete-recreate issues, and safer DB recreation/eviction flows.
• Skill & tools refinements: scoping of skill ingestion and agentic skills under access control, safer name lookups, and new models and flows to support skill improvement proposals and runs.
• Search & logging: optional search history logging toggle and improved prompt templates for global context summarization.

Bug Fixes

• Fixed a class access error and several skill ingestion edge cases (guarding name lookup and scoping adds under access control).
• Resolved issues where delete would unintentionally recreate databases.
• FastEmbed embedding format problems were addressed, plus added real mock and retry allowances for more stable tests and ingestion.
• Fixed slow eviction of subprocesses and ensured caches are evicted on Queue exit to free resources promptly.
• Prevented subprocess watchdog from killing workers when parent is PID 1 and added more robust subprocess lifecycle handling.
• Fixed LLM/model adapter issues (including llama-cpp / Ollama related fixes) and updated deprecated model usages.
• Addressed WAL file corruption cases for Ladybug graph DB.

Technical Changes

• Large internal refactor and many tests: added adapter caching CI workflow, expanded unit and performance tests for caching and global-context flows.
• Refactored LRU/cache code, added closing LRU cache utilities, and set clearer limits (20GB used during development, but configurable via env).
• Added a number of new modules and files for global context indexing, skill improvement, engine models, and retrieval (agentic retriever).
• Refactored subprocess harness and session handling to cover new init failure paths and to harden multiprocessing behavior.
• Updated package-locks, pyproject and poetry lock; removed/changed certain optional dependencies for cross-platform CI and MCP builds.

Generated by Cognee Release Notes Generator on 2026-05-16