Stored memories have paths you can inspect
OpenViking gives retained context a viking:// address. Its README describes a virtual filesystem where developers can inspect and edit documents, memories, and skills. Consider it when seeing what an agent retained would help you operate that agent. Our sandbox run established installation and build results, but did not establish working retrieval.
The README separates directory abstracts, longer overviews, and full source content into loading levels. An agent can read a summary before opening the underlying file. That gives operators material to inspect, while leaving them responsible for checking whether generated summaries and retained memories represent their sources accurately.
What happened when we ran it
Our sandbox installed OpenViking commit 92285e5 in 60 seconds, recorded 35 dependencies using 38 MB on disk, and completed the build in 6 seconds. Tests stopped after 2 seconds with a missing pytest_asyncio import. The command exited 4 while loading tests/conftest.py, before the suite could exercise retrieval.
The supplied traceback identifies line 11 of conftest.py and ends with ModuleNotFoundError for pytest_asyncio. It does not show why the module was missing, whether a documented development configuration would supply it, or whether application tests would pass after it became available.
The 2026-10-13 run recorded 3 CPUs, 8 GB RAM, and image ghcr.io/astral-sh/uv:python3.12-bookworm. The image label does not verify actual runtime versions. The checkout had 5,069 files, about 1,115,732 source lines, and occupied 103.5 MB. The run recorded 30 CI workflow files, a Dockerfile, Compose configuration, and a tests directory. Pip-audit reported 0 known vulnerabilities. That audit does not measure retrieval quality or establish service security.
These measurements concern commit 92285e5. The current documentation and the later plugin report below are separate evidence, not behavior verified on that checkout.
Models, access, and licensing affect adoption
The configuration guide uses ~/.openviking/ov.conf and configures embeddings separately from the vision-language model. API-based providers need appropriate credentials. The README also documents local Ollama. The configuration guide describes a default local embedding model, with a separate inference dependency and a model download on first use unless cached. These are documented setup options; our lab did not verify model calls or document processing.
Changing ov.conf requires a restart. The guide distinguishes startup configuration from a limited set of runtime settings. Choose the models and persistent workspace before treating a package build as a usable memory system.
For authenticated agent integrations, the README requires a user key rather than a root key for reading and writing memories. It also directs operators to configure authentication before exposing the server beyond localhost. Check the setup path for your intended client, including the installer's platform requirements.
The README identifies the main project as AGPLv3. It lists Apache 2.0 for crates/ov_cli and examples, while the Hermes plugin retains MIT licensing and framework adapters declare AGPLv3. Check the component you intend to use or modify against your licensing requirements.
Search can return duplicates beyond the requested limit
The retrieval guide describes global vector retrieval with optional reranking. Direct find requests do not use session context. Search can use session context and an LLM for intent analysis when enabled. Target directories constrain search scope, but retrieval does not navigate directories recursively.
For search, the limit applies to each planned query. Merged results may exceed that limit, and deduplication across queries is not guaranteed. An application with a fixed context allowance should inspect the returned set before passing it to a model.
Start a pilot with a document whose answer you know and another document that should be irrelevant. Check the returned source paths, then compare direct retrieval with a session-aware request. Inspect result counts and duplicates as well as relevance. These are proposed acceptance checks, not experiments completed in our sandbox.
Inspect Claude Code captures before routine use
Issue 6314, opened October 19, 2026, reports that the Claude Code memory plugin turns API-request records into conversation turns. The reporter lists OpenViking 0.5.1, plugin 0.7.12, Windows, and Claude Code 2.1.295 or later. The reported duplicates contain system prompts and injected context. This plugin report is separate from our Python test-startup failure.
An October 20 comment points to a proposed shared-parser fix and reports validation of a synthetic request-envelope case. The commenter explicitly says this does not verify the reporter's Windows transcript or a live capture. The comment also says existing capture cursors and previously archived duplicates are not migrated. Inspect a small captured archive before connecting routine sessions, and check that every retained field belongs there.
The supplied repository metadata records a push at 2026-10-20T04:45:12Z, observed on October 20, 2026. That dated activity and the issue discussion do not establish overall maintenance health or resolution of this capture report.
Compare deployment and storage choices
LangMem provides functional memory primitives and native integration with LangGraph's storage layer. Consider it for an existing LangGraph application when using that storage integration is the priority.
Mem0 presents library, self-hosted server, and cloud options. Compare those deployment paths with the service you would operate for OpenViking. No comparative retrieval measurements are supplied for either alternative.
OpenViking merits a narrow pilot when browsing retained context would help your team operate an agent. Before connecting a daily workflow, retrieve a known document and inspect a captured session archive. The supplied build result supports starting that evaluation; the blocked tests leave application behavior unresolved.

