Release Notes - v1.2.0
Release Date: 2026-06-21
Changes: v1.2.0.dev1 → main
Summary
Cognee 1.2.0 focuses on smarter session distillation and visualization, tighter security defaults, and improved robustness across LLM, storage and vector backends. This release adds new API surfaces (proposals + inline skill ingest), major session refactors for reliability and performance, and several important bug fixes and config clarifications to make deployments safer and easier to operate.
Highlights
- Smart session distillation: auto-distill during improve(), distilled nodes marked in visualization, and a reworked distillation pipeline for reliability and fewer duplicate documents.
- New proposals API and inline skill ingestion to streamline skill workflows and proposal retrieval.
- Security and config changes: public registration is disabled by default and backend access control/auth posture has been clarified; many previously undocumented managed settings are now in the .env.template.
- Robustness fixes: better handling for LLM/embedding edge-cases (retry caps, over-length inputs, fallback endpoints), improved S3 IAM handling and LanceDB Windows long-path support.
- Major internal refactor: session management split into smaller components (session_turn, session_embeddings, distillation modules) for clearer flows and easier debugging.
Breaking Changes
- Environment variable renames for completion/embedding token knobs: LLM_MAX_TOKENS -> LLM_MAX_COMPLETION_TOKENS and EMBEDDING_MAX_TOKENS -> EMBEDDING_MAX_COMPLETION_TOKENS. Update your .env or deployment configs if you relied on the old names.
- Security posture changes: public registration is disabled by default and global settings are more restricted. ENABLE_BACKEND_ACCESS_CONTROL now implies API auth and per-user/dataset DB isolation — single-user deployments that call visualize_graph() without a dataset may need to set ENABLE_BACKEND_ACCESS_CONTROL=false or adjust auth settings. Review the updated .env.template and CLAUDE.md sections for how auth flags interact.
- Pinned/patched dependency changes: instructor version constrained (< 1.15.3) and other dependency lower bounds raised for security; ensure your pinned deployments are compatible.
New Features
- New Proposals endpoint (Get Proposal) and inline skill ingest API — easier programmatic skill proposal and ingestion workflows.
- Session: auto-distill support in improve() — Cognee can now trigger automatic distillation as part of session improvement and mark distilled nodes in visualization for easier inspection.
- New/expanded endpoints for remember, forget, and improve flows (expanded API routers and handlers).
- Example demos added for session distillation and stepwise session flows to help you try the new features locally.
Improvements
- Session distillation reworked: batched curator, per-lesson writers, reuse of stored embeddings for novelty search, and emission of one document per lesson — reduces duplication and improves distillation quality.
- Visualization: dataset context is resolved earlier and distilled nodes are marked in the graph; front-end memory_map view and supporting templates updated for a clearer memory graph UX.
- LLM & embedding resiliency: retry capping for LLM calls, handling of over-length embedding inputs, and support for fallback completion endpoints in the OpenAI adapter.
- S3FileStorage now falls back to boto3 credential chain (IAM role support) when explicit credentials are not provided.
- LanceDB lifecycle improvements: automatic long-path prefixing on Windows and safer prune() path checks.
- CLI/tests & docs: many tests stabilized and expanded; .env.template now documents additional managed settings (LLM tuning, chunking, session cache, rate limits, graph and vector DB tuning).
- Swagger/OpenAPI: most endpoints now render correctly in Swagger UI.
Bug Fixes
- Fixed multiple Swagger rendering issues so more endpoints show properly in the API docs.
- Fixed UI Docker support for Colima.
- Fixed instructor dependency constraint and other dependency bumps for security patches.
- Fixed various skill ingestion and lookup issues that caused missing or incorrect skill data to be stored.
- Fixed session QA vector handling: missing QA vector indices are treated as empty, QA recall vectors are kept in sync and cleaned up when needed.
- Fixed offloading blocking filesystem IO in ontology API to avoid blocking the event loop.
- Fixed search type-check bug that could wrap datasets into lists incorrectly.
- Fixed memory visualization initialization so the dataset DB context is resolved before graph engine startup.
- Fixed several LanceDB and cache adapter bugs and added unit tests to prevent regressions.
Technical Changes
- Large internal refactor of session code: SessionManager was slimmed down and turn processing was extracted into session_turn; session_embeddings and multiple distillation modules were added and reorganized for clarity and testability.
- Many new unit and integration tests added across sessions, visualization, adapters, and storage backends to improve reliability.
- Adapters: Ollama client made async, OpenAI adapter now supports fallback endpoints, and many LLM adapters received tweaks to support string responses and direct request bodies.
- Blocking filesystem operations in some API paths were offloaded to worker threads to avoid blocking FastAPI's event loop.
- Numerous internal formatting, linting, and CI workflow tweaks (nightly/e2e adjustments).
Generated by Cognee Release Notes Generator on 2026-06-21