中/EN
Open Source
Open Source

OpenViking: A Context Database for Agents, Open-Sourced by ByteDance at 33,000 Stars

ByteDance's Volcano Engine open-sources OpenViking: a context database that unifies agent memories, resources, and skills into a viking:// virtual file system you browse with ls/tree/find instead of a black-box vector store. Core designs include L0/L1/L2 tiered loading (saves tokens), recursive directory retrieval, observable retrieval traces, and session-committed memory. 33,172 stars on GitHub (API snapshot 2026-08-25), #6 on the weekly trending chart with +3,540 stars. Official benchmarks: Claude Code long-conversation memory jumps 57.21% to 80.32%, input tokens down 34.3%-91.0%. AGPL-3.0 - mind the viral clause for commercial SaaS.

Published August 25, 20268 min read
<!-- openviking-resource | open-source | OpenViking: A Context Database for Agents, Open-Sourced by ByteDance at 33,000 Stars -->

The default recipe in the agent world for the past two years: want to give your model memory? Bolt on a vector store - chunk documents, embed them, dump them in, and recall top-k snippets into the prompt at query time. The problem is that this path is a black box for developers. Where results come from, why these snippets, how to debug a bad retrieval - none of it is visible. ByteDance's Volcano Engine has open-sourced volcengine/OpenViking with a different paradigm: make the agent's context a database - one you can browse with ls, tree, and find as a virtual file system.

GitHub API snapshot (2026-08-25): 33,172 stars / 2,527 forks, AGPL-3.0, Python, created 2026-01-05, with a push landing today (actively maintained), ranked #6 on this week's GitHub trending chart with 3,540 stars gained - 33,000 stars in eight months makes it one of the fastest-climbing agent infrastructure projects of the year.

Scope note: star counts and repo status are API snapshots (2026-08-25); this is a representative teardown based on the official README and docs, not a long-term hands-on; benchmark figures come from the official report.

1. What It Solves: From Black-Box Vector Store to File System

OpenViking positions itself as an open-source AI Agent context database: it unifies everything an agent needs at runtime - memories, resources, and skills - into a single virtual file system mounted under its own viking:// protocol. Every memory, resource, and skill gets a URI, and the agent locates its context deterministically the way a developer works with files, rather than throwing a query into a black box and hoping.

The difference in perspective is fundamental. In the vector-store path, context is a pile of embeddings reachable only by semantic search; in the file-system path, context is something you can browse, traverse, address exactly, and audit. When retrieval goes wrong, you can open the directory tree and see which path it actually took - a qualitative change for anyone debugging agents.

One sentence to distinguish it from our Mem0/Zep/Letta memory-tools comparison: those are memory-layer components that sit beside an existing agent doing extraction and retrieval; OpenViking is a unified context database that manages memories, resources, and skills in one file-system namespace - a bigger ambition by one layer.

2. Core Design I: the viking:// Virtual File System and L0/L1/L2 Tiered Loading

The whole system hangs off one URI namespace: viking://memories/, viking://resources/, and viking://skills/ each own a slice. The agent gets file-system semantics: list a directory, view a tree, reference one specific memory - all deterministic, no retrieval luck involved.

The more consequential piece is tiered processing on write: as content enters OpenViking it is rendered into three layers - L0 summary, L1 overview, L2 detail. Shallow tasks load only L0/L1; deep dives descend to L2. This turns the most expensive problem in context engineering - the token budget - into a structural design: load by depth as needed instead of stuffing the window. The 34.3%-91.0% input-token reduction in the official benchmark owes mostly to this tiering.

3. Core Design II: Recursive Directory Retrieval, Observable Retrieval, Session-Committed Memory

On the retrieval side, OpenViking also skips the plain vector top-k route. Three designs worth singling out:

  • Recursive directory retrieval: vector search first locates the highest-scoring directory, then drills down layer by layer to the target. Results arrive with their directory context attached, not as isolated text fragments.
  • Observable retrieval: every query keeps its full directory-browsing trace. When a result is wrong, you can see which path, which branch, produced it - the single biggest debugging pain of black-box vector stores, removed outright.
  • Session-committed memory: when a session ends, commit asynchronously extracts user preferences and accumulated agent experience into long-term memory. By the next session, they are already waiting under viking://memories/.

4. Benchmarks: Not an Incremental Bump

The official 0.3.22 benchmark report (Doubao 2.0 Pro as VLM) posts aggressive numbers:

  • LoCoMo long-conversation user memory: OpenClaw improves from a native 24.20% to 82.08%; Hermes from 33.38% to 82.86%; Claude Code from 57.21% to 80.32%. All three frameworks land in the 80%+ band after integration.
  • Cost and latency: input tokens down 34.3%-91.0%, query latency down 58.45%-66.10%.
  • tau2-bench task success: experience memory adds +6.87pp on retail and +11.87pp on airline.

Standard caveat: this is the vendor's own benchmark - trust the direction, re-test the magnitude. But the shape of the result, "all frameworks pulled level to 80%+ after plugging it in," indicates the gain comes from the infrastructure layer rather than any framework-specific tuning.

5. Up and Running in Three Minutes

Python 3.10+. One pip install, three commands to start the server, with the ov client bundled:

bash
pip install openviking --upgrade
openviking-server init      # interactive wizard: configure providers into ~/.openviking/ov.conf
openviking-server doctor    # self-check
openviking-server           # start

# ov client CLI (bundled):
ov add-resource https://github.com/volcengine/OpenViking
ov ls viking://resources/
ov tree viking://resources/volcengine -L 2
ov find "what is openviking"

Providers include Volcengine, OpenAI, Codex OAuth, Kimi, GLM, and local Ollama (which can auto-install the runtime and pull models) - you are not locked to ByteDance's stack. One gotcha: add-resource without --wait runs asynchronous indexing - query immediately after adding and you may find nothing yet, because semantic processing takes time.

6. Ecosystem: It Catches Nearly Every Major Agent Framework

OpenViking ships integrations for Claude Code, Codex, OpenClaw, Hermes, Cursor, TRAE, OpenCode, and pi, works with generic MCP clients, and has a LangChain/LangGraph connector (injected recall + automatic session-memory commits). Three companions round it out: OpenViking Helper (beta, macOS/Windows x64 desktop console for visually configuring agent integrations, parsing session traces, and managing local memories and skills), VikingBot (an agent framework built on OpenViking, pip install "openviking[bot]"), and the playground at openviking.ai/studio. Those 33,000 stars rest partly on each framework's community contributing its own integration docs.

7. License and Fit: You Must Clear the AGPL-3.0 Gate First

The license is the one line to read before committing: AGPL-3.0 is viral for commercial redistribution, and SaaS scenarios need particular care - if you serve users over a network using a modified OpenViking, you must release your derivative code. If that stings, two doors exist: Volcano Engine's Managed SaaS (personal tier, 50 files free to try) or the Self-Managed commercial edition (deployed in your own environment, offline isolation supported). The good news: the open-source edition is not crippled - under AGPLv3 there is no feature gate.

Who it's for: heavy agent builders (long conversations, long-term memory, multi-session accumulation), individuals who want persistent context for Claude Code/Cursor-style coding agents, and teams whose infrastructure can live with AGPL. If you just want to swap the vector store in an existing RAG pipeline - that is not what this does.

One-line closer: when an agent's context goes from a vector black box you cannot inspect to a directory tree you can ls, memory becomes debuggable, auditable, and tier-billable for the first time.

References

This article is based on the official README and docs (as of 2026-08-25); star counts are API snapshots and benchmark figures come from the official report, subject to the official source.

This article is AI-assisted and human-edited. Last updated: 2026-08-25

Related

Open Source

DeepInfer Open-Sourced: 155T Tokens/Day, Engine or Factory?

On October 20, 2026 DeepSeek announced full open sourcing of DeepInfer, its production-grade inference framework - the first time the official production engine is available instead of open-source parts plus guesswork. The repository deepseek-ai/DeepInfer ships under Apache-2.0 and measured 17,902 stars, 1,836 forks, and 188 open issues on launch day, while the companion token codec deepseek-recipe is MIT-licensed; the two repositories carry different licenses. Every performance figure below is vendor-reported and has no third-party reproduction yet: a single cluster serves 155 trillion tokens per day (about 38 dollars per card per hour at off-peak pricing), first-token latency of 400 milliseconds, 230 tokens per second generation, 96 percent prefix-cache hit rate, and Prefill throughput per card rising from 60,000 to 1.5 million tokens per second when caching counts. SWA Bounded Replay pins the recompute of cache misses to the last 128 tokens, which nearly halves total Prefill compute and makes cost independent of context length; DSpark speculative decoding drafts up to 5 tokens, adds another 6 percent to live throughput, and yields 14K tokens per card per second; MoE expert replicas reshuffle every 10,000 steps so the busiest card runs only 9 percent above average load. Hardware covers both NVIDIA SM100 and Huawei Ascend 950 NPUs, and setup means a Python 3.12 virtual environment, the DEEPINFER_ACCELERATOR variable, and a source build that pulls DeepSeek-V4.1-Flash-1017 weights by default. The boundaries matter: DeepSeek states that Prefill and Decode disaggregation, data-parallel load balancing, and the KVCache management layer are still not fully open because they couple too deeply to internal infrastructure - this is the engine, not the whole factory - and no official benchmark against vLLM, SGLang, or TensorRT-LLM exists. In production, 8-card servers handle Prefill while 64 to 128-card supernodes handle Decode, with KVCache kept beyond 72 hours to avoid recomputation.

Oct 22, 202610 min read
Open Source

AI Memory Framework, Open-Sourced: Tencent's 3B Tops Opus 5

Tencent Hunyuan open-sourced all three layers of Hy Memory 2.0 on 2026-10-16: the model layer Hy-Memory-30B-A3B (30B total parameters with about 3B active per token, a sparse MoE on HuggingFace at tencent/Hy-Memory-30B-A3B, arXiv 2610.22196), the SDK layer hy-memory (2.0.3 on PyPI at writing time) and the tool layer MemHub (CLI v1.0.3, bootstrapped via the npm memhub-cli package). License check: the GitHub API reports NOASSERTION (non-standard text), but the raw LICENSE file is Apache-2.0 with a Tencent preamble and no added restrictions - this article follows the LICENSE text. Official documents mix two spellings, "Hy-Memory-30A3B" and "Hy-Memory-30B-A3B"; we follow the HF repository name. Official benchmarks (vendor-reported): first place at 55.49 in a four-way memory-system comparison, ahead of Opus 5 (54.62) and DeepSeek-Flash (53.64), while Qwen3-30B-A3B scores only 44.37; MemoryArena 55.31 percent, leading runner-up EverOS by 10.2 points; CodeMemoryBench 64.2/35.6; SWE-ContextBench 28.3; LoCoMo 72.6 percent with memory tokens cut from 1159 to 703 (-39.3 percent). Internal business validation (vendor-reported): stock-picking multi-hop recall up from 50 to 90 percent, iWiki up 48.6 percent relative. Positioning is a memory-writer workhorse (analyst take, secondary): the model specializes in writing and updating memory and can in principle sit under existing systems like Mem0 or MemOS. MemHub tool count in two versions: the README lists 5 while the official WeChat article says 10 - go by actual CLI output. Five limitations stated as-is: vendor-only benchmarks without third-party replication, preview-stage API churn, MemHub limited to Linux/macOS, a 30B-class model that is nontrivial to self-host, and thin English docs. Repository: github.com/Tencent-Hunyuan/Hy-Memory (verified reachable). Division note: ai-agent-long-term-memory-sop (2026-08-08) covers the build-your-own mem0/Letta route; this piece covers Tencent's ready-made stack.

Oct 20, 202610 min read
Open Source

AI Agents Can Now Reverse Engineer Apps: 84k Stars in a Week

morluto/rea (Reverse Engineer Anything) topped GitHub Trending: verified via GitHub API on October 11, 2026 - 84,088 stars, MIT, TypeScript; the star-growth chain cites three sources separately (~10,400 via trending data, 47,884 via a third party quoting the GitHub API, 84,088 verified by us). Repo: github.com/morluto/rea. It is an MCP server plus CLI (npx rea-agents setup) that wires your coding agent (Claude Code/Codex/Cursor/Gemini CLI and more) into a reverse-engineering toolbox: native binaries through Hopper/Ghidra/IDA, static JS/Electron with no engine needed, .NET assemblies, APKs (JADX), firmware (Binwalk/Unblob), EVM bytecode, HAR/mitmproxy captures and websites. v6.0 (release notes, October 8, 2026) added pwntools offline ELF layout, crash inspection, Windows PE x86 and LLDB observation, with v6.1-v6.3 following within 24 hours. Honest limits: it does not recover original source, does not clone apps wholesale, and every conclusion ships with evidence and stated caveats; analysis runs locally so the target binary never leaves your machine. README-cited showcases: DX-Ball function rebuilt in C passing 3,205 original x86 cases with all 63 bytes reproduced; the Notion clipboard bridge traced end to end. created_at 2026-04-14 (API) vs a reported October 7 launch - both stated, no exact date asserted.

Oct 11, 202610 min read