EverOS changelog: what's new each month
Every stable EverOS release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 5 months covered; the current month updates daily.
October 2026 · so far
1 release: 1.4.2
October brought performance improvements and fixes for multi-project setups, with changes to embedding concurrency defaults and vector database handling.
Action needed
- Milvus collections must be rebuilt once after upgrading to this version.
Highlights
- Embedding concurrency default increased from 5 to 10 for better performance.
- Interactive search no longer blocked by background cascade operations.
- Fixed memory overwrites when using the same owner across multiple projects.
- Resolved file descriptor leaks in long-running servers with LanceDB.
Improved
- Cascade now optimizes LanceDB tables at most once per minute instead of continuously.
- Response IDs scoped to project level for better multi-project isolation.
- LanceDB indexing behavior improved for multi-project roots created before this release.
Fixed
- File descriptor leak on LanceDB index rebuilds in long-running servers.
- Memory conflicts when the same owner's data exists in multiple projects.
- Interactive search responsiveness degraded by cascade embedding operations.
September 2026
4 releases: 1.3.0 → 1.4.1
September brought major stability improvements and expanded platform support, with a new Windows release, better vector search capabilities, and numerous fixes to backend operations and data consistency.
Action needed
- everos cascade sync, cascade fix --apply and cascade rebuild now refuse operation under certain conditions.
- [cascade] scanintervalseconds changed from a hardcoded value to a configurable setting.
Highlights
- Native Windows support now available with MSVC runtime included.
- Vector (ANN) indexes added for improved performance on vector columns.
- LLM-guided multi-round episode retrieval enables more sophisticated memory access patterns.
- Concurrent writes on Milvus backend no longer lose data.
- Multiple backend stability fixes for LanceDB and Milvus.
New
- Optional Milvus derived-index backend support.
- LLM-guided multi-round episode retrieval.
- Independent decider configuration via new [decider] section.
- Unified benchmark harness covering LoCoMo and other adapters.
- Cascade snapshot control with quiesce and rebuild commands.
- Configurable extraction LLM transport via [llm] section.
- Native Windows support with MSVC runtime.
- Vector (ANN) indexes for vector columns.
Improved
- Vector search now uses cosine distance on LanceDB.
- Milvus datetime round-trips are exact below millisecond threshold.
- Milvus queries stay within engine result window.
- SQLite pool saturation now fails visibly with configuration details.
- Cascade scan interval is now configurable (default 30 seconds).
- Watcher paths relativised more efficiently without per-event resolve calls.
- OME strategy attempts bounded to 1,800-second timeout.
- Permanent embedding request errors no longer retry on HTTP 400, 401, 403.
Fixed
- POST /api/v1/memory/add no longer fails on freshly resolved installs.
- Concurrent updates on Milvus backend no longer lose writes.
- Watcher late modified events for deleted paths now recorded correctly.
- PyArrow floor raised to 25.0.1 to fix timestamp column materialisation.
- OpenAI dependency pinned below 3.x to resolve fresh install compatibility.
- POST /memory/add now properly rejects role="tool" messages without content.
August 2026
2 releases: 1.2.2 → 1.2.3
August brought stability improvements to EverOS's database layer and fixes to agent skill handling. The month focused on robustness, with better error detection, bounded resource usage, and hardened security.
Highlights
- Health check endpoint now provides detailed cascade readiness information with human-readable reasons and internal counters.
- New cascade rebuild command allows recovery from corrupted or drifted indexes by clearing the queue and re-indexing from scratch.
- All write-lock critical sections are now bounded with deadlines, preventing unbounded resource consumption.
- Vector dimension mismatches now fail immediately with clear error messages instead of opaque database errors.
- Path-traversal hardening added to LLM-generated agent skill names to prevent security vulnerabilities.
New
- Cascade rebuild CLI command for index recovery
- Prune operation for LanceDB maintenance with per-table write lock
- Query vector dimension validation at configuration time
- OME strategy retry backoff between attempts
- Detailed health check cascade readiness metrics
Improved
- Startup schema verification now detects column type drift, not just missing or extra columns
- LanceDB maintenance split into lock-free compaction and cleanup operations
- Benign LanceDB commit conflicts no longer count as failures
- Exception logging no longer renders frame locals, reducing log volume on async stacks
- Agent skill extraction no longer stuck in retry-then-dead-letter loops
- Agentic search on agent memory uses skill-shaped rerank
- POST /api/v2/ome/trigger no longer masks strategy state
Fixed
- Index corruption caused by column type drift detected at startup
- Unbounded index growth from unfinished LanceDB maintenance
- Renamed skills leaving orphan directories that pollute the system
- Opaque database errors from vector dimension mismatches
July 2026
6 releases: 1.1.1 → 1.2.1
July brought major API improvements and infrastructure updates to EverOS, including native OpenTelemetry tracing, the new v2 API endpoint, and flexible dependency configuration. The month also saw performance optimizations for search and numerous stability fixes.
Action needed
- All business endpoints now served under /api/v2; /api/v1 retained for backward compatibility.
- MemoryRoot.default() renamed to MemoryRoot.resolve(); default() kept as alias.
- ProviderNotConfiguredError now returns HTTP 422 CAPABILITYUNAVAILABLE instead of previous behavior.
Highlights
- Launched /api/v2 endpoints for all business operations, aligning the open-source API with EverOS Cloud.
- Added native OpenTelemetry tracing for memory operations, exportable to any OTLP backend like Langfuse.
- Made embedding and rerank providers soft dependencies so EverOS runs with only an LLM configured.
- Improved hybrid search performance through heap-driven lazy expansion and global top-N competition.
- Fixed critical concurrent Knowledge cascade upsert issues and LanceDB index bloating problems.
New
- LoCoMo benchmark runner with TOML configuration and staged execution pipeline.
- everos cascade backfill CLI command for three-phase memory upgrade (vectors, clusters, skills).
- LanceDB schema v2 supporting nullable vectors for rows without embeddings.
- Langfuse OpenTelemetry integration example for operation tracing.
- GET /health endpoint returning typed HealthResponse with capabilities and disabled features.
- Startup banner warning for unbackfilled memory rows with NULL vectors.
- PyPI Trusted Publishing workflow using OIDC for secure releases.
Improved
- Hybrid search with heap-driven lazy expansion achieving 91% stable LoCoMo targets.
- Agentic search performance reaching 93% LoCoMo targets.
- CI coverage expanded to Python 3.12 and 3.13.
- OpenTelemetry synthetic child spans limited to responses with stage details for cleaner traces.
- Profile middleware now properly re-raises inner handler exceptions instead of swallowing them.
- OpenAPI application version kept in sync with package metadata.
- Provider error handling returns HTTP 422 with section-aware hints pointing to everos.toml.
Fixed
- Concurrent Knowledge cascade upserts now use atomic INSERT ... ON CONFLICT DO UPDATE.
- LanceDB FTS withposition crashes during optimize() fixed; now defaults to False.
- Agent-track search broken by deprecated IS NULL filter.
- Stale indexed rows and permanently failed queue items from disappeared source files.
- Consecutive optimize() failures now escalate from warnings to errors instead of silent failures.
June 2026
3 releases: 1.0.0 → 1.1.0
June brought EverOS 1.0.0 stable release with a complete memory architecture based on plain Markdown files, plus two follow-up releases adding Knowledge management, Reflection capabilities, and security hardening.
Highlights
- Markdown files serve as the primary memory store, readable and editable with any text editor or Git.
- Hybrid retrieval combines BM25, vector, and scalar filtering in a single query.
- Security hardening to prevent path-traversal attacks in file operations.
- New Knowledge APIs for document creation, storage, and management.
- Reflection orchestration for periodic episode merging and refinement.
New
- Knowledge APIs and storage for creating, listing, and patching documents.
- Reflection orchestration system for merging and refining episodes.
- Event IDs and run-record storage in the OME runtime.
- Typed exception handlers and consistent error handling across APIs.
- Multimodal usage guide and error semantics.
- Pluggable OpenAI-compatible providers for LLM, embedding, and reranking.
Improved
- Search and get behavior for agent-owned memory.
- Path-traversal defense with character whitelisting on caller-supplied identifiers.
- Write containment in MarkdownWriter to prevent files outside the memory root.
- Documentation and OpenAPI schema coverage for new capabilities.
- Test coverage expanded across new features.
- OME runtime support with additional tracking capabilities.
Fixed
- Accuracy drift found during documentation audit.
- Multimodal error semantics corrected after end-to-end verification.
- Path-safety guard now applied consistently across senderid, appid, and projectid.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. EverOS: pricing, features and alternatives · All changelogs