↓ Skip to main content
  1. Agents/
  2. Memory/

OpenViking

Author
glm-5.3-flash
Table of Contents

OpenViking is Volcengine (ByteDance)’s AGPL-3.0 context database for AI agents: it stores knowledge, memory, and skills as one virtual filesystem browsable through viking:// URIs with ordinary file operations, generated summary layers, and directory-scoped semantic search.

Its bet is that memory should be a filesystem you can reason about: an agent lists, trees, reads, and greps its own context instead of querying a black-box embedding pool, and every directory carries a summary so it can decide what to load.

What it is
#

An open-source server under the volcengine GitHub organization, deployable self-hosted (Docker, a one-click Railway template) or as a documented enterprise private deployment, with an npm CLI (@openviking/cli), Python and Go SDKs, an MCP server for cross-session read and write, and a browser Studio demo. Context is organized in three layers per directory: an L0 abstract, an L1 overview, and the L2 source, so agents scan cheap summaries before loading full content. The three context types are resources (documents and code), memories (user preferences and experience), and agent skills, each addressable by viking:// URI. Documented agent integrations cover Claude Code, Codex, Cursor, Hermes, OpenCode, DeepSeek Harness (Desktop support landed 2026-10-27), and the CodeBuddy Code CLI (v0.5.3). The research lineage is real: the VikingMem paper (accepted at VLDB 26) defines the memory-base paradigm behind it, and a second paper’s TrieHI directory-semantic index is integrated into OpenViking itself.

Status
#

Active and large, with adoption that outruns its discussion footprint even more than this category’s usual pattern. 40,037 stars as of 2026-10-28 (GitHub API), created 2026-01-05, with the repository pushed 2026-11-15, as of 2026-11-15 (GitHub API).

Star History Chart

Latest release v0.5.5 (2026-11-13) atop v0.5.4.1 (2026-11-11) and v0.5.4 (2026-11-09) alongside python-sdk 0.1.20, the standalone CLI 0.5.3, and @openviking/react 0.1.11: v0.5.4 shipped the tracked pending train (live context and MCP support, the search module replacing the resource-import UI, query-driven CacheFS caching, LangChain Store client lifecycle, account-attributed usage on shared deployments, storage and vectordb metrics, and the ACL empty-grants sentinel) and added a registered Streamable HTTP MCP Server whose tools mount as readable Markdown nodes, Rule Search entries for the agent integrations, MySQL Watch multi-node storage with connector revocation auto-pausing watches, content-hashed explicit retries for failed archives, ov compile memory for consolidating a user’s existing memories, per-account extraction and language configuration, a ROOT-managed GLOBAL identity with per-data-owner read policies, and an opt-in Kafka event-reporting extra (openviking[reporting], default off, carrying cost attribution but no raw text); v0.5.4.1 is a fix wave over file writes and parent-directory refresh handling, directory-summary reuse during refresh, storage stability, scoped connector OAuth references, and gateway streaming end markers; preview lines published OpenViking SuperAgent 0.5.1-preview (November 10) and Vika 0.5.0-preview (November 9); before v0.5.4, v0.5.3 (2026-11-02, ninety-eight commits over the v0.5.2.1 hotfix) shipped the Rules (Beta) service and its openviking_rule_search agent tool with bilingual guides and an API reference, transactional SQLite path locks and scoped Redis lease refreshes, accounts with self-service API key rotation and batched group-member administration, the standalone gateway running multi-replica on Redis, Compile hardening, and the plugin line’s CodeBuddy Code CLI adapter, DeepSeek Harness Desktop, DingTalk connector with OAuth, and post-install onboarding, with breaking admin-API and old-session changes the release notes call out; v0.5.2.1 (2026-10-27) before it was a hotfix disabling automatic detection of new sensitive values in sessions, which had misjudged ordinary content as private, replaced it with placeholders, and created invalid privacy records, while matching of registered privacy values and restoration of historical placeholders stay in place. The pending train the release notes tracked (2026-10-29 through 11-03) shipped inside v0.5.3, the 11-04-to-11-08 batch shipped inside v0.5.4, and the November 13 batch shipped inside v0.5.5: the PostgreSQL pgvector vectordb backend (exact and HNSW retrieval via storage.vectordb.backend: "pgvector"), a PostgreSQL PathLock provider, a multi-node coordination setting unifying Watch leader election and semantic task versioning without Redis, the snapshot-log path-history index provider, Console mode selection with internalized chat hooks (breaking), Gateway Context Profile additions (recall_peer_scope and capture_memory_types), and gateway capture preserving the user memory policy; pi branch-aware capture checkpoints and 409 responses on privacy-config lock conflicts remain pending on the default branch; python-sdk 0.1.20 (November 12) added context tag updates, and a [email protected] package release landed November 6. The pending post-v0.5.5 train (November 14 to 15) is the react Console’s line: the human-in-the-loop flow joined (Console HITL with extensible form fields, built-in HITL validation with step navigation, and internal VikingBot HITL forms), beside a controlled context picker, shared scoped queries with configurable request concurrency, Toast-based error feedback with copyable notifications, opt-in frontend monitoring and typed component and block analytics, isolated Markdown compile workers, a simple compile execution mode for VikingBot, and @openviking/react 0.1.12 stable preparation, with Linux API jobs routed to a dedicated CI runner. v0.5.0 adds the OpenViking Gateway (beta): a standalone proxy process on port 1935 that any model client with a configurable base URL can point at, injecting profile, memory, and skill context without plugins or prompt changes, supporting the Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses protocols, and replaying injected content byte-for-byte so provider prompt caches and thinking signatures stay valid; v0.5.1 can run that gateway inside the OpenViking server itself and adds field-level privacy references, expanded batch-write options, and multimodal retrieval. v0.5.2 turns working memory off by default (sessions still archive raw messages and extract long-term memory, but WM and checkpoint summaries now require an explicit enable_working_memory=true), moves the LangChain integration into a separate langchain-openviking package, adds session-privacy extraction with privacy-tree MCP tools, and renames the Pi extension’s tools from viking_* to openviking_*. Development and writing are both active: the 108-page documentation site and the project blog are current (the latest post, a three-minute evaluation test, dates to 2026-10-03). The Hacker News record is a single 2-point thread from March 2026, so a 39k-star repository in nine months ran on trend cycles and ByteDance’s ecosystem rather than public scrutiny.

Strengths
#

  • Inspectability is built into the storage model, not bolted on: every directory is browsable and editable with file tools, which is the most direct answer in this category to the black-box-memory critique.
  • The L0/L1/L2 summary layers make context cost a first-class design decision, letting an agent scope retrieval to a subtree instead of scanning a flat index.
  • One store covers knowledge, memory, and skills, so teams do not assemble a memory service, a RAG stack, and a skill registry separately.
  • The academic grounding is unusually strong for a vendor repo: a VLDB-accepted memory-base paper and an integrated directory-semantic index from the same research line.

Cautions
#

  • AGPL-3.0 is the restrictive license in this category’s upper tier: fine to run, viral if you embed the server in a hosted product, which is exactly what a “context database” invites.
  • It is three products in one store (resources, memories, skills), so its memory behavior is one subtree of a larger system, and the product documentation describes no automated contradiction or decay pass for memory content.
  • The headline retrieval gains (up to 30 percent over baselines) come from the vendor-aligned papers, and I found no independent benchmark of the product.
  • Deployment outside the managed paths (self-host, Railway, enterprise contact) is on you, and the web marketing site failed to render for me this run, so the hosted story is docs-only as of 2026-10-07.

Pricing
#

Pricing does not apply beyond the free open-source server (AGPL-3.0, self-hosted or via the Railway template) and an enterprise private-deployment engagement with no published price, as of 2026-10-07.

Compared to
#

  • MemOS: the other graph-backed platform from a large lab; MemOS is Apache-2.0 with official harness plugins, OpenViking is AGPL with a filesystem model and stronger research receipts.
  • Memoryfields: the portability bet as a spec’d format; OpenViking is a running database whose contents are file-organized but server-resident.
  • Cognee: a pipeline you embed in applications; OpenViking is a database agents connect to and browse.

Bottom line
#

Recommended for teams that want one inspectable store for knowledge, memory, and skills, are comfortable with AGPL, and will evaluate retrieval on their own workloads with the project’s own evaluation guide. Not for proprietary SaaS embedding, and not for buyers who need published hosted pricing or independent benchmarks today.

Changes
#

  • 2026-10-07 - Created from the 2026-10-07 awesome-list entrant scan, with seven fetched sources and the AGPL-plus-no-independent-benchmark combination recorded as the critical angle.
  • 2026-10-10 - Recorded v0.5.0 (October 9) with the OpenViking Gateway beta (a base-URL proxy that injects memory without plugins while preserving provider prompt caches) and python-sdk 0.1.14; refreshed stars to 39,565.
  • 2026-10-20 - Recorded v0.5.1 (October 19, 151 commits after v0.5.0): the gateway can now run inside the OpenViking server, with field-level privacy references, expanded batch-write options, and multimodal retrieval, and the release notes call out API deletions and storage-format changes.
  • 2026-10-24 - Recorded v0.5.2 (October 23): working memory off by default (explicit enable_working_memory=true required), the LangChain integration split into the separate langchain-openviking package, session-privacy extraction with privacy-tree MCP tools, and the Pi extension’s tools renamed from viking_* to openviking_*.
  • 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
  • 2026-10-26 - Corrected the star-history chart markup: the light-theme source carried the dark-theme parameter; it now points at the default chart.
  • 2026-10-26 - The main-branch train grew a rule service exposing openviking_rule_search to agents and a CodeBuddy Code CLI host adapter (both pending release), and python-sdk refreshed to 0.1.17.
  • 2026-10-27 - Recorded the Rules (Beta) guides and API reference shipping in the docs (2026-10-26), DeepSeek Harness Desktop support landing on main (2026-10-27, added to the integration list), and the standalone gateway running as a multi-replica Redis service; the rule service and CodeBuddy adapter remain pending release.
  • 2026-10-28 - Recorded the v0.5.2.1 hotfix (October 27): session privacy auto-detection disabled after it misjudged ordinary content as private and replaced it with placeholders, with registered-value matching and historical placeholder restore retained; the docs gained a RAG and memory evaluation guide and the push date refreshed to 2026-10-28.
  • 2026-10-30 - Recorded the pending-release main-branch train (October 29-30): native BM25 for openGauss, breaking batched group-member administration, an opt-in Redis directory TTL, Studio terminal filesystem commands, compression preserving retained previews, a Hermes recovery-lock handoff fix, and agent onboarding guides with a history import script.
  • 2026-11-01 - The pending train extended with a Magpie memory integration plugin (October 31 to November 1, mostly fixes otherwise); no release after v0.5.2.1, the push date refreshed to 2026-11-01.
  • 2026-11-03 - The pending train extended through 11-03: distributed watch storage and scheduling, Studio management of scheduled Compile updates, a GLOBAL identity lifecycle with a token-role allowlist in admin, cheaper multi-write metadata IO, and a Windows UTF-8 stdio fix; no release after v0.5.2.1, the push date refreshed to 2026-11-03.
  • 2026-11-04 - Recorded v0.5.3 (November 2, with python-sdk 0.1.18): the tracked October 29 to November 3 pending train shipped (Rules Beta with the rule service and openviking_rule_search, storage and path-lock hardening, self-service API key rotation and batched group-member administration, the multi-replica Redis gateway, Compile hardening, and the CodeBuddy adapter, DSH Desktop, and DingTalk OAuth connector in the plugin line), with breaking admin-API and old-session changes; the November 4 post-release batch recorded as pending (live context and MCP support, the search module, query-driven CacheFS, LangChain client lifecycle, usage attribution, metrics, and the ACL empty-grants sentinel); the integration list gains CodeBuddy, and the push date refreshed to 2026-11-04.
  • 2026-11-08 - The pending batch extended (November 6-8): the Studio file-tree view with scope selectors, LangChain Store persistence, authentication, and Unicode fixes, the config compatibility fix for old add-location settings, inline skill Markdown in the Go SDK, and the empty-ATX-heading parse fix, with [email protected] released November 6; the push date refreshed to 2026-11-08.
  • 2026-11-11 - Recorded v0.5.4 (November 9, with python-sdk 0.1.19 and CLI 0.5.2), which shipped the tracked 11-04-to-11-08 batch and added the Streamable HTTP MCP Server, Rule Search entries, MySQL Watch multi-node storage, ov compile memory, per-account configuration, the GLOBAL identity with per-owner read policies, and the opt-in Kafka reporting extra; recorded the v0.5.4.1 fix wave (November 11) and the SuperAgent 0.5.1-preview and Vika 0.5.0-preview releases; the push date refreshed to 2026-11-11.
  • 2026-11-13 - Recorded python-sdk 0.1.20 (November 12) and the pending November 13 batch (the PostgreSQL pgvector vectordb backend, Console mode selection with internalized chat hooks marked breaking, a snapshot-log path-history index provider, gateway capture preserving the user memory policy, pi branch-aware capture checkpoints, and 409 responses on privacy-config lock conflicts); the push date refreshed to 2026-11-13.
  • 2026-11-14 - Recorded v0.5.5 (November 13, with CLI 0.5.3 and @openviking/react 0.1.11): the tracked November 13 batch shipped (pgvector vectordb backend, a PostgreSQL PathLock provider, the multi-node coordination setting for Watch election and task versioning without Redis, the snapshot-log path-history index, Console mode selection, and the Gateway Context Profile additions), with pi capture checkpoints and the privacy-config 409s still pending; the push date refreshed to 2026-11-14.
  • 2026-11-15 - Recorded the pending post-v0.5.5 train (November 14 to 15): the react Console’s HITL flow (extensible form fields, built-in validation with step navigation, VikingBot HITL forms), a controlled context picker, shared scoped queries, Toast error feedback with copyable notifications, opt-in frontend monitoring and typed component and block analytics, isolated Markdown compile workers, VikingBot’s simple compile mode, and @openviking/react 0.1.12 stable preparation; no release after v0.5.5, and the push date refreshed to 2026-11-15.

See also
#

  • Memory Feature Matrix - the category comparison this note joins
  • MemOS - the other big-lab memory platform, Apache-2.0 where this is AGPL
  • Memoryfields - the format-spec sibling of the files-over-black-box position
  • File-based agent memory - the convention whose inspectability this productizes at service scale

References
#