v1.5.3.dev1 — Integrations, safety checks, and search improvements

Release Date: 2026-08-26
Changes: v1.5.3 → v1.5.3.dev1

Pull Requests: #3161, #3274, #3453, #3496, #3536, #3715, #3884, #4132, #4188, #4222, #4440, #4465, #4487, #4519, #4525, #4528, #4531, #4534, #4560, #4561, #4563, #4567, #4569, #4573, #4575, #4576, #4577, #4578, #4585, #4587, #4588, #4593, #4596, #4598, #4603, #4606, #4607, #4608, #4609, #4610, #4611, #4612, #4619, #4620, #4627, #4634, #4635, #4640, #4642, #4644, #4645, #4646, #4647, #4653, #4663, #4664, #4666, #4667, #4669, #4670, #4671, #4676

Summary

This development release focuses on new integrations (Linear and GitHub App), better developer tooling (config preflight and automated docstring sync), and search/recall improvements for more reliable results. It also includes several security hardenings (WebSocket guards, telemetry fixes), performance work on ingestion, and internal refactors that may affect Postgres-backed deployments.

Highlights

• Linear agent integration and GitHub App organization connector — connect tasks and repositories directly to Cognee.
• New config 'preflight' and doctor CLI command — proactively checks your config and suggests fixes before you run services.
• Default hybrid search — search now uses both keyword and vector matching for more relevant results out of the box.
• Telemetric insights GitHub Action and scripts — daily automated extraction and prompts to surface telemetry insights.
• Security hardening for real-time subscriptions — run-subscription WebSockets are now authenticated and guarded.

Breaking Changes

• Postgres graph adapter rewrite and relocation: The Postgres graph adapter has been substantially rewritten and files were reorganized (moved to a postgres_demo path). What it means: self-hosted deployments that use the older Postgres adapter should review their database adapter configuration and custom code that imports the old Postgres adapter. You may need to update import paths or adapter settings before upgrading.
• Removal of MCP workspace UI: The MCP workspace user interface components were removed. What it means: if you relied on the previous MCP workspace UI bundle or its front-end files, plan to migrate to the updated MCP server or alternative management tooling.

New Features

• Linear agent integration: Adds a built-in agent that can create and update Linear issues from Cognee. What it does: lets you automate ticket creation and track tasks in Linear directly from Cognee flows. Why it matters: integrates your memory-driven workflows with your issue tracker so actions taken in Cognee can create or update tickets without manual work.
• GitHub App organization connector: A connector that lets Cognee access organization-level GitHub data. What it does: enables ingestion and synchronization of repositories and organization events when you install the GitHub App. Why it matters: makes it easier to bring code, PRs, and issue context into Cognee for search, recall, and automation.
• Config preflight and doctor CLI command: A new command-line check that validates your configuration and suggests fixes before you start services. What it does: runs a set of environment and configuration checks and reports actionable problems. Why it matters: reduces setup errors and speeds troubleshooting for new and existing deployments.
• Telemetry-insights GitHub Action and extraction scripts: An automated workflow that daily extracts telemetry metrics and runs a prompt to surface notable insights. What it does: collects and summarizes operational metrics each day and stores the result in the repo workflows. Why it matters: helps teams spot usage trends and anomalies without building custom telemetry pipelines.
• Self-correcting router docstring workflow: An automated workflow that keeps router function documentation (docstrings) in sync with code. What it does: checks and can suggest or apply fixes to function docstrings used by the routing layer (the small functions that decide which tool or skill to call). Why it matters: ensures the human-readable descriptions used to decide routing remain accurate, improving tool selection and developer experience.
• Brains-summary and dataset WebSocket endpoints: New real-time endpoints that stream summary updates about 'brains' (a brain is a Cognee workspace or knowledge collection) and datasets (the collections of documents you've added). What it does: provides live updates when dataset summaries change. Why it matters: enables dashboards and UIs to show near-real-time state without constant polling.
• Expose operation records in activity feed: The activity feed now includes operation records. What it does: surfaces a history of operations (adds, deletes, pipeline runs) in the activity UI and APIs. Why it matters: gives teams visibility into what actions ran and when, simplifying audits and debugging.

Improvements

• Default hybrid search: Search now defaults to 'hybrid' mode that combines vector similarity (semantic search) and keyword matching. What it does: returns results that balance exact keyword matches and semantic relevance. Why it matters: improves relevance for general queries without extra configuration.
• Per-user preference personalization: User-level preferences are now applied to personalize results. What it does: stores and applies user settings that can influence ranking and result presentation. Why it matters: helps tailor search and recall outcomes to individual users.
• Per-plugin agent identity provisioning: Agents (automations) can get unique identities per plugin. What it does: allows each automation or plugin to operate with its own identity and credentials. Why it matters: improves traceability and security when multiple agents act on behalf of users.
• Docker Sandboxes example and starter kit: A new example showing how to run sandboxed memory kits in Docker. What it does: provides a ready-to-run example for experimenting with memory sandboxes. Why it matters: lowers the barrier for testing Cognee features locally and safely.
• Delete session info on document deletion: When you delete documents, related session data is removed. What it does: prunes session-level references to deleted documents. Why it matters: reduces stale or confusing data in sessions after content is removed.
• Recall warm-up short-circuit: Recall now supports a fast warm-up path to improve initial latency for certain queries. What it does: avoids expensive initialization on the first recall in many cases. Why it matters: reduces first-call delays for recall-heavy workloads.
• Multiple LLM integration fixes: Includes correct handling of local temperature settings, rate-limit defaults for stage-routed configs, and better model qualification for local adapters. What it does: fixes inconsistencies in how model config values are applied to local LLM servers. Why it matters: ensures generated outputs match your configured settings and avoids rate-limit surprises.

Performance

• Reduced S3 calls during ingestion: The file add() path now makes fewer S3 requests per file (down from ~13 to 4 in typical cases). Practical impact: faster and cheaper ingestion of large numbers of files from S3.
• Reproducible nightly corpora and corpus builder tooling: Added tools and a documented process for rebuilding the nightly test corpora. Practical impact: more reliable performance testing and easier reproduction of benchmark datasets.
• CI reliability fixes and reduced nightly burn: Several CI improvements reduce flaky tests and unnecessary long-running nightly jobs. Practical impact: faster CI feedback and fewer wasted build minutes.

Security

• Authenticated run-subscription WebSockets: The real-time 'run-subscription' WebSocket endpoint is now guarded and negotiates authentication before sending sensitive data. What it does: requires valid credentials for clients subscribing to pipeline or run events. Why it matters: prevents unauthenticated clients from receiving potentially sensitive pipeline events.
• Fix for telemetry leak in Recall: A recall-related fix reduces inadvertent telemetry exposure during follow-up queries. What it does: removes extraneous telemetry information from certain recall flows. Why it matters: reduces accidental leaking of telemetry into logs or responses.
• Signature verification for integrations: GitHub and Linear integration handlers include verification of inbound webhooks. What it does: checks signatures to ensure events originate from the expected provider. Why it matters: prevents forged webhook events from being processed.

Bug Fixes

• Require query on recall and search endpoints: Recall and search endpoints now validate that a query is present and return actionable errors when missing.
• Fix: Ollama env-var check now judges resolved LLM configuration (no false negatives on model config).
• Fix: decode_token_list now decodes each token id correctly.
• Fix: Strip markdown fences from litellm_native JSON fallback outputs so downstream parsers are not confused.
• Fixes to multiple CI flakes, Windows and Python 3.10 unit test breaks.
• Fix: Resolve several Ladybug Cypher syntax incompatibilities used by graph traversal helpers.

Technical Changes

• Large refactor of graph and database layers, including a rewrite of the Postgres-backed graph adapter and new 'unified' store plumbing.
• New and expanded API surface: many v1 endpoints for remember, recall, search, improve, visualize, sessions, and activity were added or expanded.
• Session, operation recording and provenance improvements: added richer operation records, pipeline run models, and provenance delete planning.
• Developer CI and workflow additions: new GitHub Actions and scripts for telemetry insights, router docstring sync, release automation, and nightly/performance reporting.
• Infrastructure and LLM plumbing: many small adapters added for local LLMs, embedding engines, and tokenizer logic to better support different providers and local fallbacks.

Dependency Updates

Updated:

• cryptography: >=43.0.0,<51 → >=43.0.0,<50

Compatibility

— The Cognee Team · 2026-08-26