volcengine/OpenViking
A database that lets AI agents browse their own memory and files like a file system
OpenViking is an open-source context database that unifies an AI agent's memories, resources, and skills under one virtual filesystem addressed as viking://. It pre-processes every entry into three loading tiers - a one-sentence abstract, an overview, and full details - and loads only as much as a task needs, cutting token usage. Reported benchmarks show large gains in accuracy and latency compared to agents' native memory.
What it does
- Memories, documents, and skills are each given a viking:// address so agents can browse and search them with familiar commands like ls, tree, and find instead of querying an opaque vector store
- On write, content is processed into L0 (a roughly 100-token one-sentence abstract), L1 (a roughly 2k-token overview), and L2 (full original content); retrieval first locates the best-matching directory via vector search, then drills down layer by layer
- Every retrieval keeps a record of the directory-browsing path it took, so a wrong result can be traced back to exactly how it was produced
- After a session ends, OpenViking asynchronously extracts user preferences and agent experience into long-term memory
- On the LoCoMo long-conversation memory benchmark, adding OpenViking raised accuracy for OpenClaw, Hermes, and Claude Code from 24-57% (native memory) to 80-83%, while cutting input tokens by 34.3-91.0%; on tau2-bench multi-turn agent tasks, task success rose by 6.87 percentage points (retail) and 11.87 points (airline) over the same LLM without memory
Why it matters
When AI agents handle long histories and large document sets, re-reading everything or relying on opaque vector search drives up cost and errors; OpenViking makes context retrieval predictable and debuggable, like working with files. It plugs directly into tools like Claude Code, Cursor, and LangChain, so it can be tried in real agent workflows right away.
Terms in this repo
- viking:// protocol · the addressing scheme OpenViking uses for memories, resources, and skills, treated like file paths
- L0/L1/L2 · three loading tiers - one-line abstract, overview, full details - loaded only as deep as needed
- LoCoMo · a benchmark measuring how well a system retains user memory across long conversations
- tau2-bench · a benchmark measuring an agent's success rate on multi-turn conversational tasks
- VikingBot · an AI agent framework built on top of OpenViking
Repository description (English)
Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Open on GitHubTrending repos
- cathrynlavery/diagram-designA skill that makes AI coding tools draw magazine-quality diagrams instead of generic rounded boxes
- public-apis/public-apisA giant crowd-curated directory of free APIs for developers
- semantica-agi/semanticaAn open-source graph infrastructure that lets AI agents show their work, not just their answers
- cactus-compute/needleA 14MB AI model small enough to run tool-calling on a phone or watch, without internet
- unslothai/unslothA desktop app that lets you run and train AI models on your own computer, no coding required
- macro-inc/macroAn all-in-one workspace where email, chat, docs, tasks, and CRM are cross-linked and share one AI memory
- harry0703/MoneyPrinterTurboAn open-source tool that turns a single topic or keyword into a finished short video, complete with script, footage, subtitles, and music
- basecamp/omarchyA ready-made, opinionated Linux setup built by DHH
Latest from METAL LAB
- Grok leaks conversations, location data via commands hidden in ciphertext
- Cursor lets agents handle long-running tasks with new "/goal" command
- NVIDIA releases Switchyard, an LLM routing proxy for coding agents
- Caveman Cuts Claude Code Token Usage by 33% Using Caveman-Speak
- Apple Keeps 2027 Target for Camera AirPods, Qwen3.8-27B Makes 80 Tool Calls Locally