Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

OpenViking changelog: what's new each month

Every stable OpenViking release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 7 months covered; the current month updates daily.

November 2026 · so far

8 releases: 0.5.3 → 0.1.12

November brought major infrastructure improvements to OpenViking, including new Rules integration for agents, PostgreSQL deployment options, and significant enhancements to storage, authentication, and real-time context handling. The month also delivered extensive UI refinements and new connector integrations.

Action needed
  • Old session paths (viking://session/...) now return INVALID_URI; use new path format viking://user/{user_id}/sessions/... exclusively.
  • Admin user key listing now shows only the caller's own keys.
  • Remote PathGlob queries controlled by storage.vectordb.enable_path_glob setting.
  • Preset global policy initialization requires explicit server.user_config_defaults.preset_global_policy enablement.

Highlights

  • Rules service now available in beta, exposing a rule search tool to agents for querying applicable rules before execution.
  • PostgreSQL support added as a deployment option with pgvector vector backend and PostgreSQL PathLock for multi-host deployments.
  • MCP and real-time context integration enables streaming HTTP MCP servers with tool retrieval and mounting as readable nodes.
  • Streamlined session management with automatic restoration on failure and new memory organization tools.
  • Separated file writing from directory refresh to prevent slow directory processing from blocking file operations.

New

  • Rules (Beta) service with openviking_rule_search tool for agents and comprehensive API documentation.
  • PostgreSQL pgvector backend support for vector storage with precise and HNSW retrieval options.
  • MCP Server registration for streaming HTTP servers with tool mounting and synchronization tracking.
  • MySQL Watch storage backend with connector authorization auto-pause and recovery.
  • Snapshot path history indexing using derived indexes for efficient commit lookup.
  • Static soul.md support as editable Agent instructions read with priority by VikingBot.
  • Account-level configuration for extraction toggles, generation language, and credential parsing.
  • Batch context tag setting API and context usage counters with recency tracking.

Improved

  • User self-service API Key rotation and role-based access control with case-insensitive input.
  • Group members now added in batch, with improved administrator visibility and role downgrade support.
  • Gateway session expiration configurable per second and tools null handling in responses.
  • Compile retrieval with optional rerank decoupling, image source support, and L2 chunk sizing.
  • Multi-node concurrent operations improved by limiting locks to actual target ranges only.
  • Directory stat and list operations now count only visible L2 records, excluding L0/L1 summaries.
  • Console mode selection for conversations, commands, or both with independent routing.
  • HITL question parsing with ordered answers, field validators, and per-question custom renderers.

Fixed

  • PathLock premature expiration from recursive deletion and Chinese path legacy lock recovery.
  • Vectordb Collection lifecycle and startup issues with incomplete metadata writes.
  • File processing coordination with parent-directory refresh maintaining independent indexing.
  • Gateway streaming error end markers and native import preservation of Connector Watches.
  • Directory semantic overview sampling stability with proper L2 chunk configuration.
  • OIDC account rejection for invalid identifiers instead of character-based mapping.

October 2026

7 releases: 0.4.23 → 0.5.2.1

October brought major architectural improvements with the new Gateway serving as a unified interface between clients and model services, plus significant enhancements to memory injection, retrieval, and multi-instance deployment capabilities.

Action needed
  • Working Memory now disabled by default in sessions; use enable_working_memory=true to restore previous behavior.
  • Minimum SDK dependency raised to 0.1.17 for Compile support.
  • LangChain import path changed: replace openviking.integrations.langchain with langchain_openviking (install langchain-openviking package).
  • Gateway must be enabled in ov.conf with gateway.enabled: true before startup; independent deployment still supported.

Highlights

  • Gateway (version 0.5.0) now acts as a dedicated process between clients and model services, handling memory injection and conversation compression automatically.
  • Memory injection and retrieval system redesigned to preserve prompt cache validity and attach relevant memories directly to user messages.
  • Retrieval refactored with unified global vector recall and reranking, removing directory recursion and parent score propagation.
  • Privacy configuration now supports field-level references and atomic updates with scanning and repair capabilities.
  • Multi-instance deployment enhanced with optional Redis-based session idle auto-commit and new MySQL/Redis+MySQL Account Store providers.

New

  • Gateway process (openviking-gateway on port 1935) as independent deployable component with Anthropic Messages and OpenAI Chat Completions support.
  • Memory injection prepends user profile and memories/skills directory to sessions, with byte-by-byte replay to preserve provider caching.
  • Automatic conversation compression when approaching model context window limits (default 90%).
  • Context profile management in Studio Gateway page with tool round/time/token limits.
  • Session privacy auto-extraction, recovery, and privacy tree MCP tools.
  • Batch write processing with per-file tags, ACL, and content version conflict detection.
  • Search mode for CLI with context output assembly and support for multi-image retrieval.
  • Incremental compile updates with scheduling and version history.

Improved

  • Retrieval now uses single global vector recall per query with unified reranking, applies scoring threshold post-rerank.
  • find/search commands accept events_time_decay_protection parameter for time-decay sorting.
  • ls/tree commands return has_more pagination status and support directory filtering with L0/L1 display options.
  • MCP tool declarations now include behavior annotations and return isError flag on call failure.
  • Permission management expanded with list_users, list_groups, get_acl, set_acl tools and ACL parameters for write/add_resource.
  • Collector filtering unified across all agent plugins (pi, Codex, DSH, OpenCode, Claude Code) with clean-filter-truncate pipeline.
  • Installation process now downloads packages from docs site, prompts language selection, and installs ov CLI by default.
  • CLI rebranded with E brand logo and solid colors; documentation reorganized with agent-first quick start guide.

Fixed

  • Session commit returning 400 error for GPT-6 and later OpenAI models treated as reasoning models.
  • Gemini async client creation per request, resolving 'attached to a different loop' errors.
  • Windows + uv launcher readiness status misdetection causing 900-second service wait timeout.
  • Subsequent session commits after initial commit now properly index generated memories for retrieval.
  • Directory retrieval blocking new files in subdirectories.
  • LangChain/LangGraph mapping, text chunking, async initialization, and health status detection.

September 2026

6 releases: 0.4.18 → 0.0.4

September brought substantial improvements to resource management, retrieval capabilities, and system stability. Major additions include resource-level access controls, transactional file operations, pagination and sorting, and runtime configuration management.

Action needed
  • GET/PATCH /api/v1/admin/accounts/{accountid}/settings and GET/PUT /api/v1/admin/agent-evolution marked deprecated; use configuration API instead.
  • ACL must be explicitly enabled with account configuration {"acl":{"enabled":true}}; disabled by default for backward compatibility.
  • Docker images no longer include ripgrep; RAGFS no longer probes or calls external rg command.

Highlights

  • Resource-level ACL system with user and group authorization, supporting read/write/manage permissions with hierarchical inheritance.
  • Transactional copy and move operations that migrate vector records and handle file overwrites and directory merging.
  • Pagination and sorting for file listing and tree operations across all APIs and SDKs.
  • Tag-based file system filtering integrated into listing, search, and grep operations.
  • Account-level runtime configuration allowing VLM, embedding, and other settings to be changed without restarting.

New

  • Resource-level ACL with user, group, and account-level authorization granularity.
  • POST /api/v1/fs/cp endpoint and ov cp command for transactional file copying with vector record migration.
  • Pagination (offset/limit) and sorting for ls and tree commands.
  • Tag writing on first file write and tag-based AND filtering across operations.
  • Account and cluster-level runtime configuration API.
  • Skill-level retrieval returning best match per skill and skill directory injection into sessions.
  • MCP addskill tool and OpenCode v2 plugin support.
  • Feishu mindmap parser and external OAuth token management for Watch tasks.

Improved

  • File copy and move now use fine-grained path locking (file-level Exact, directory-level Tree).
  • Vector retrieval automatically filters results by effective ACL permissions.
  • Watch tasks support pause state, recent execution history, and Connector incremental refresh.
  • Filter operations can now work independently of query, returning results from metadata without computing similarity scores.
  • VectorDB local schema updates now atomic and non-destructive, appending fields without rebuilding indexes.
  • HTTP connection pooling reused across VikingDB and Volcengine clients.
  • Query embedding reused across context assembly and skill retrieval within single requests.
  • Git Watch automatically disables on authentication failure; improved MCP proxy connection management.

Fixed

  • File system scanning error type corrections.
  • MCP proxy timeouts now reported as -32004 instead of unreachable.
  • camelCase isError and toolResult fields now properly recognized in tool results.
  • Session batch Phase 2 extraction aligned with auto-commit policy.
  • cp/mv index reading, empty resource imports, path locking, queue isolation, and local vector storage process locks.
  • MCP grep error reporting and search parameter handling for list mode.

August 2026

7 releases: 0.4.12 → 0.4.17.1

August brought significant enterprise capabilities and cross-platform alignment to OpenViking, including multi-modal audio and video understanding, enterprise authentication, and SDK feature parity across Python, Go, and TypeScript.

Action needed
  • Direct Sandbox is disabled by default for Compile; Skill operations fail with SKILLCAPABILITYUNAVAILABLE if required binaries or environment variables are unavailable.
  • Unbounded xxhash dependency in v0.4.14 could cause vector persistence failures; users may need to reindex affected content after upgrading to v0.4.15.

Highlights

  • Agents can now directly read and write Viking URIs with tree, write, and edit operations via MCP.
  • Server-side context assembly in search mode reduces client complexity by handling recall planning, token budgeting, and deduplication.
  • Multi-modal audio and video understanding based on Volcano Ark backend with long upload and result caching support.
  • Cross-language SDK alignment brings feature parity for find, search, recall, resources, sessions, and management across Python, Go, and TypeScript.
  • VikingBot can now discover, cache, and use Skills hosted in remote OpenViking instances.

New

  • OIDC and LDAP authentication for enterprise deployments.
  • QueueFS support for Redis single node, Cluster, and Sentinel modes.
  • Portable Agent Plugins 1.0 package with stdio-to-HTTP MCP proxy.
  • DeepSeek Harness memory plugin with automatic recall and session capture.
  • DeerFlow and TRAE integration guides with MemoryManager and MCP options.
  • MCP native media return for images and audio content blocks.
  • Private TOS URL import with temporary credentials via args.tossignature or args.tosaccess.
  • Viking URI shorthand viking://~ for current user's root directory.

Improved

  • Session Auto Commit v2 with token count, message count, and idle time triggers plus session-level policies.
  • Resource import speed and observability with background task processing when wait=false.
  • Agent Evolution with experience sources, trajectory lineage, result aggregation, and snapshot refinement.
  • Recall Context services unified as V3 with dynamic memory extraction policies per user.
  • Memory extraction policies now configurable per user by administrators.
  • Directory creation now generates minimal L0 metadata and queues vectorization even without description.
  • Content write operations create missing files and parent directories for replace and append modes.
  • CLI initialization wizard expanded with VolcEngine/BytePlus tier options and custom interaction configuration.

Fixed

  • VectorDB write errors now correctly propagate through the pipeline instead of silently failing.
  • Compile lifecycle handling for rapid consecutive commits now maintains clearer causal relationships.
  • HTTP keep-alive configuration and Redis startup recovery mechanism added.
  • File path normalization and URI validation on server-side processing.
  • AnyDoc document model adapted to version 0.2.
  • Resource import with private Git authentication fixed for both Watch and addresource operations.

July 2026

6 releases: 0.4.7 → 0.4.11

July brought major infrastructure improvements to vector search, file handling, and integrations across the OpenViking ecosystem. The month focused on performance optimization, expanded IDE support, and strengthening the memory and training framework.

Action needed
  • Tag search changed from OR to AND semantics: multiple tags now require all to match instead of any.
  • Setting watch_interval > 0 on one-time snapshots now errors; content updates require re-upload.
  • TypeScript SDK now distinguishes cancellation (ABORTED) from timeout (DEADLINE_EXCEEDED).

Highlights

  • GPU-accelerated vector search via cuVS backend with automatic CPU fallback based on available memory.
  • Recursive web importing with Scrapy-driven crawling and configurable depth, page limits, and path filtering.
  • New integrations for Cursor, TRAE, and OpenCode with lifecycle hooks for automatic context recall and capture.
  • Memory v3 extraction pipeline now always active with streaming updates and session training support.
  • Tag search semantics changed to strict AND matching across all SDKs and APIs for consistent results.

New

  • cuVS GPU vector search backend supporting brute-force exact and CAGRA approximate retrieval with optional auto-enable.
  • Recursive web import driven by Scrapy with depth, max_pages, path inclusion/exclusion, and external link controls.
  • Cursor integration with lifecycle hooks, MCP, rules, and memory skills for automatic context management.
  • TRAE and TRAE CN integrations supporting multiple lifecycle hooks and immediate memory submission.
  • External Connector integration for delegated resource imports with background task status polling.
  • prune_orphans reindex mode to identify and remove orphaned vector records with dry-run preview.
  • OpenViking Helper for macOS beta for local agent detection and session tracking.
  • Context takeover for pi extension using OpenViking archive summaries when token threshold is reached.

Improved

  • Reindex now uses configured VLM concurrency to respect global concurrency controls.
  • Reindex writes switch to content owner context to prevent admin tasks from changing resource ownership.
  • MCP compact description mode reduces token overhead from tool descriptions.
  • Local file uploads via signed temp_upload tokens automatically indexed.
  • Scrapy-based web import switched to trafilatura for body extraction with sitemap/RSS paths preserved.
  • Directory listing now prioritizes directories then sorts by name.
  • Filesystem glob operations pushed to Rust backend with pagination support.
  • Streaming memory updater groups concurrent submissions by peer and memory type with patch-merge application.

Fixed

  • Resource auto-naming no longer swallows lock contention errors during concurrent imports.
  • VLM requests now omit unset default max_tokens.
  • Embedding auth errors no longer infinitely requeue.
  • Consistency check rejects file:// URIs to prevent local path misuse.
  • Subdirectory overview summaries now batch-processed to reduce duplication.
  • Locking issues in QueueFS sidecar and image rewrite paths resolved.

June 2026

8 releases: 0.3.23 → 0.4.6

June brought major architectural changes to OpenViking's identity and storage systems, alongside new collaboration features in Web Studio and significant improvements to the CLI experience. The release cycle introduced breaking changes to support a modern User/Peer context model while maintaining backward compatibility for legacy data.

Action needed
  • viking://agent/... paths are no longer write targets; new memory, resource, session, and skill data must use viking://user/... paths.
  • agent_id is now legacy-only and maps to request-level actor_peer_id; do not configure both simultaneously or the server will reject the request.
  • Legacy role_id memory isolation is no longer supported; use the new User/Peer model to express isolation boundaries.
  • Legacy agent_id clients should no longer explicitly pass message peer_id; the system handles this automatically.

Highlights

  • CLI config management was completely restructured with interactive configuration editing, language selection, and improved diagnostic views through the new ov status command.
  • Web Studio now includes an integrated Playground sidebar for viewing context trees, running terminal operations, and managing connections with account/user/agent identity controls.
  • Storage architecture shifted from agent-centric to user-centric namespacing with the new viking://user/... path, with legacy viking://agent/... paths still readable but not for new writes.
  • VikingFS gained Git-backed snapshot capabilities supporting commit, log, show, and restore operations across HTTP API, Python SDK, and Rust CLI.
  • Web Studio added user-visible metrics including token trends, context commits, and heatmaps, plus improved resource import with sitemap, RSS, and Atom feed support.

New

  • Interactive CLI config wizard with ability to add, edit, delete, and switch configurations through ov config commands.
  • Language selection for CLI display via ov language / ov lang command with automatic prompting on first use.
  • Web Studio Playground sidebar showing context trees, terminal operations, and Agent panel interactions within the same workspace.
  • Connection & Identity management page supporting account/user/agent creation, API key generation and rotation, and connection saving.
  • VikingFS Git-backed snapshot management with commit, log, show, restore operations backed by local .ovgit or S3/TOS-compatible storage.
  • Whole-site resource import supporting sitemap, sitemap index, RSS 2.0, and Atom feeds with automatic resource tree generation.
  • VikingDB BM25 Grep engine with auto mode keyword recall capability while preserving local-only search mode.
  • Go HTTP SDK alongside Python HTTP SDK, with code navigation endpoints for opencode-plugin support.

Improved

  • Resource watch configuration no longer requires explicit 'to' parameter when watch_interval is set, with automatic binding to stable root_uri.
  • Bot experience recall is now configurable for single-task and evaluation scenarios with new recall_exp_first_round_only, exp_recall_limit, and exp_recall_max_chars settings.
  • CJK-aware token estimation extended to Python and plugin sides, reducing budget underestimation for Chinese, Japanese, and Korean sessions.
  • Claude Code and OpenClaw plugins now write structured tool parts to OpenViking instead of inline text-only tool calls and results.
  • CLI and SDK surfaces improved with clearer validation errors, safer request field emission for older servers, and new Go SDK set_tags support.
  • Session archive failures are now skippable and OpenClaw auto-commit behavior was tuned for better reliability.
  • Storage and session behavior hardened including queuefs semantic processing for non-directory memory URIs, glob URI scheme preservation, and event-summary fallback.
  • add_resource enhanced with improved parent/create_parent semantics and CLI --parent-auto-create flag.

Fixed

  • Auth role serialization now consistently serializes request identity roles as strings across storage, resource, session, queue, watch, summarizer, and semantic-processing paths.
  • Session ID encoding issues were resolved to prevent session identification errors.
  • Legacy task records are now properly ignored during RAGFS shape probe operations.
  • Storage and session behavior now correctly handles non-directory memory URIs in queuefs semantic processing.
  • Server listening addresses like 0.0.0.0 and :: are now properly mapped to loopback client URLs to prevent connectivity issues.
  • Feishu URL host matching was tightened to improve parsing accuracy.

May 2026

5 releases: 0.3.18 → 0.3.22

May brought significant improvements to OpenViking's web interface, memory handling, and API performance. The Web Studio replaced the legacy console, batch operations reduced HTTP overhead, and resource processing became more robust across file types and URL detection.

Action needed
  • Legacy console removed; Web Studio is now the primary interface at /studio.
  • Console API /api/v1/console/* endpoints now require timezone parameter for date/hour buckets; clients must pass IANA timezone or receive server default.
  • Usage and audit SQLite schema incompatible with old format; existing tables deleted and rebuilt on first startup, with potential data loss for short-retention records.
  • OpenClaw plugin no longer registers memory_search tool; use ov_search instead.

Highlights

  • Web Studio now serves as the primary console interface, packaged with Docker and pip distributions, with OAuth setup integrated.
  • New batch message API endpoints allow writing up to 100 session messages in a single request, reducing latency for multi-turn imports.
  • Memory embedding now uses configurable templates that combine key fields with content for better semantic recall.
  • HTTP profiling support added for request-level performance debugging via the CLI and API.
  • Document conversion moved to background threads to prevent blocking the async event loop on large files.

New

  • Web Studio frontend console workspace with PWA support and mobile optimization.
  • Batch session message writing via POST /api/v1/sessions/{session_id}/messages/batch and CLI command.
  • Request-level HTTP profiling with server.profile_enabled switch and profile=1 query parameter.
  • Memory embedding_template schema field to replace field-level searchable marks.
  • Trajectory retrieval_anchor and derived_from StoredLink for improved memory retrieval and replay.
  • ov_search tool as the unified search interface, replacing memory_search conflicts.
  • NVIDIA NIM routing support in LiteLLM VLM backend.
  • Lightweight query planner config for intent analysis in search.

Improved

  • Web Studio resource reader enhancements and memory file preview showing raw content correctly.
  • OAuth authorize UI migrated to Web Studio with favicon routing preserved.
  • Usage and audit timestamps now stored as UTC with timezone-aware rebucketing per request.
  • Resource semantic indexing now syncs temp source to target before DAG execution.
  • HTTP accessor improved to detect images, audio, video, Office, EPUB, and zip formats with GET fallback for unreliable HEAD responses.
  • Semantic lock handoff retries tree lock acquisition instead of triggering circuit breaker on conflicts.
  • LangChain integration batch message API usage reduces latency for multi-message writes.
  • File system and API error handling with clearer AGFS exceptions and parameter validation.

Fixed

  • CLI profile output now preserved in table, scalar, filesystem, and system command results.
  • Codex VLM backend async client cache initialization fixed to prevent runtime errors.
  • Document conversion for Word, PowerPoint, Excel, EPUB legacy formats corrected to thread-based execution.
  • PDF XObject image extraction fixed to read target objects correctly.
  • Semantic sidecars properly skipped when missing.
  • Stale OpenViking client recovery in LangChain integration.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. OpenViking: pricing, features and alternatives · All changelogs