This plugin is registered as the openviking context engine in OpenClaw.

Design Positioning

• OpenClaw still owns the agent runtime, prompt orchestration, and tool execution.
• OpenViking owns long-term memory retrieval, session archiving, archive summaries, and memory extraction.
• examples/openclaw-plugin is not a narrow "memory lookup" plugin. It is an integration layer that spans the OpenClaw lifecycle.

In the current implementation, the plugin plays four roles at once:

• context-engine: implements assemble, afterTurn, and compact
• hook layer: handles session_start, session_end, and before_reset
• tool provider: registers memory/archive tools plus OpenViking resource and skill import tools
• runtime manager: connects to and monitors a remote OpenViking service

Overall Architecture

The diagram above reflects the current implementation boundary:

• OpenClaw remains the primary runtime on the left. The plugin does not take over agent execution.
• The middle layer combines hooks, the context engine, tools, and runtime management in one plugin registration.
• All HTTP traffic goes through OpenVikingClient, which centralizes tenant headers and routing logs.
• The OpenViking service owns sessions, memories, archives, and Phase 2 extraction, with storage under viking://user/* (including viking://user/sessions/*) and viking://resources/*.

That split lets OpenClaw stay focused on reasoning and orchestration while OpenViking becomes the source of truth for long-lived context.

Identity and Routing

The plugin keeps OpenClaw session identity in session and peer metadata. It does not send an OpenViking agent identity or create an agent namespace.

The main rules are:

• reuse sessionId directly when it is already a UUID
• prefer sessionKey when deriving a stable ovSessionId
• normalize unsafe path characters, or fall back to a stable SHA-256 when needed
• peer_role=assistant is the default and writes assistant messages with peer_id=<sessionAgent>; if peer_prefix is set, the value becomes <peer_prefix>_<sessionAgent>
• peer_role=none disables peer message attribution and actor-peer routing
• peer_role=person writes user messages with peer_id derived from OpenClaw sender identity; assistant messages do not get peer_id
• data-plane recall/search/read/import/delete sends the same resolved peer identity as X-OpenViking-Actor-Peer when peer_role is assistant or person
• when OpenClaw does not provide a session agent, use its default agent main for local session and assistant peer metadata
• only add X-OpenViking-Account / X-OpenViking-User when accountId / userId are explicitly configured

This matters because OpenViking tenant identity is account/user-scoped, while OpenClaw agent identity is runtime metadata.

The recommended remote-mode configuration only needs:

• baseUrl
• apiKey
• optionally peer_role
• optionally peer_prefix when peer_role=assistant

In this setup:

• apiKey should usually be a user key
• new installs default to peer_role=assistant
• accountId / userId are advanced options only when the deployment needs explicit identity headers, such as root-key or trusted-server flows

User namespace

The plugin writes and searches user-scoped memory through viking://user/...; OpenViking resolves that alias from the request tenant and actor-peer context. Deprecated agent URI paths are not used by the plugin.

assemble Recall Flow

Auto-recall now runs through assemble(). OpenClaw calls the same context engine method in two shapes, and the plugin assigns different responsibilities to each shape:

1. Preflight assemble: params include prompt; messages is still old history. The plugin reads archive/session context back from OpenViking and rebuilds history.
2. transformContext assemble: params do not include prompt; the latest messages entry is already the current user turn. The plugin only runs long-term recall and prepends the memory block to that user message content.

During recall, the plugin:

1. Extracts query text from the latest user message.
2. Resolves the agent routing for the current sessionId/sessionKey.
3. Runs a quick availability precheck so model requests do not stall when OpenViking is unavailable.
4. Queries the configured recallTargetTypes (user,agent by default; optionally resource; use ov_archive_search and ov_archive_expand for session history).
5. Deduplicates, threshold-filters, reranks, and trims the results under a token budget.
6. Prepends the selected memories as a ## Long-term Memories section inside <openviking-context> to the current user message; it does not append a standalone synthetic user message.

The reranking logic is not pure vector-score sorting. The current implementation also considers:

• whether a result is a leaf memory with level == 2
• whether it looks like a preference memory
• whether it looks like an event memory
• lexical overlap with the current query

Session Lifecycle

Session handling is the main axis of this design. In the current implementation it covers history assembly, incremental append, asynchronous commit, and blocking compaction readback.

What assemble() does

During preflight, assemble() is not just replaying old chat history. It reads session context back from OpenViking under a token budget, then rebuilds OpenClaw-facing messages:

• latest_archive_overview becomes [Session History Summary]
• pre_archive_abstracts becomes [Archive Index]
• active session messages stay in message-block form
• assistant tool parts become toolCall (input compatible: toolUse/input is normalized to toolCall/arguments)
• tool output becomes separate toolResult
• the final message list goes through a tool-use/result pairing repair pass

That means OpenClaw sees "compressed history summary + archive index + active messages", not an ever-growing raw transcript.

What afterTurn() does

afterTurn() has a narrower job: append only the new turn into the OpenViking session.

• it slices only the newly added messages
• it keeps only user / assistant capture text
• it preserves toolCall / toolResult content in the serialized turn text
• it strips injected <openviking-context> blocks, historical <relevant-memories> blocks, and metadata noise before capture
• it appends the sanitized turn text into the OpenViking session

After that, the plugin checks pending_tokens. Once it reaches commitTokenThresholdRatio of the model context window (tokenBudget), it triggers commit(wait=false):

• archive generation and Phase 2 memory extraction continue asynchronously on the server
• the current turn is not blocked waiting for extraction
• if logFindRequests is enabled, the logs include the task id and follow-up extraction detail

This automatic path is best-effort and commit-dependent. Short but important facts can stay only in the live session until a threshold commit, /compact, or an explicit store happens.

Explicit long-term memory writes

When the user explicitly asks the agent to remember, save, or store an important long-term fact, preference, project, or decision, prefer memory_store over waiting for normal auto-capture. memory_store writes the text to an OpenViking session and calls commit(wait=true), so it is the reliable integration-side path for facts that should be available as long-term memory as soon as possible.

Use it as a complement to auto-capture, not a replacement:

• auto-capture still preserves ordinary conversation flow and batches extraction for cost and latency
• memory_store is for explicit durable-memory intent such as "remember my main project is X" or "save this preference"
• if memory_store commits but extracts 0 memories, check the OpenViking server extraction/model configuration; the explicit path triggered extraction, but the extractor did not produce a memory

What compact() does

compact() is the stricter synchronous boundary:

• it calls commit(wait=true) and blocks for completion
• when an archive exists, it re-reads latest_archive_overview
• it returns updated token estimates, the latest archive id, and summary content
• if the summary is too coarse, the model can call ov_archive_expand to reopen a specific archive

So afterTurn() is closer to "incremental append plus threshold-triggered async commit", while compact() is the explicit "wait for archive and compaction to finish" boundary.

Tools and Expandability

Beyond automatic behavior, the plugin exposes these tools directly:

• memory_recall: explicit long-term memory search
• memory_store: write explicit long-term facts into an OpenViking session and trigger commit
• memory_forget: delete by URI, or search first and remove a single strong match
• ov_archive_expand: expand a concrete archive back into raw messages
• ov_recall_trace: inspect recent recall/search trace records when traceRecall is enabled
• add_resource: import a document, directory, URL, or Git repository as an OpenViking resource when explicitly enabled
• add_skill: import or register an OpenViking skill
• ov_search: search OpenViking resources and skills, especially after importing them
• ov_read: read one exact viking:// URI returned by ov_search or ov_list
• ov_multi_read: read multiple exact viking:// URIs, useful for an overview plus sibling chunks
• ov_list: list a hit's parent directory after ov_search to recover sibling chunks, .overview.md, and related split-document context

They serve different roles:

• automatic recall covers the default case where the model does not know what to search yet
• memory_recall gives the model an explicit follow-up search path
• memory_store is for immediately persisting clearly important information when the user expresses durable-memory intent
• ov_archive_expand is the "go back to archive detail" escape hatch when summaries are not enough
• add_resource lets the agent save explicit document or repository import requests without asking the user to remember slash commands
• add_skill imports skills into OpenViking, while add_resource imports resources
• ov_search closes the loop after import by letting the user or agent confirm and consume resources and skills
• ov_read turns a ranked hit into original evidence before answering precise documentation, codebase, configuration, or procedural questions
• ov_multi_read reads overview and sibling chunks together when a split document needs more context than a single hit
• ov_list complements ov_search when a ranked hit is only one chunk of a larger procedure or document

ov_archive_expand is especially important because assemble() normally returns archive summaries and indexes, not the full raw transcript.

Resource and Skill Import

Resource and skill imports are intentionally separate because they land in different OpenViking namespaces and use different server APIs:

• resources go through /api/v1/resources and land under viking://resources/...
• skills go through /api/v1/skills and land under viking://user/skills/...

The plugin also registers explicit slash commands for manual imports:

/add-resource ./README.md --to viking://resources/openviking-readme --wait
/add-skill ./skills/install-openviking-memory --wait
/ov-search "OpenViking install" --uri viking://resources/openviking-readme
/ov-search "memory install skill" --uri viking://user/skills

Resource import supports remote URLs, Git URLs, local files, local directories, and uploaded zip files. OpenViking's built-in parsers cover common documents and media such as Markdown, text, PDF, HTML, Word, PowerPoint, Excel, EPUB, images, audio, and video. Directory imports also accept common code, documentation, and config file extensions such as .py, .js, .ts, .go, .rs, .java, .cpp, .json, .yaml, .toml, .csv, .rst, .proto, .tf, and .vue.

For HTTP safety, the plugin never sends a direct local filesystem path to the OpenViking server. Local files and directories are first uploaded through /api/v1/resources/temp_upload; directories are zipped locally with a pure JavaScript zip implementation before upload.

Runtime Mode

The plugin operates exclusively in remote mode as a pure HTTP client:

• baseUrl and optional apiKey come from plugin config
• no local subprocess is started or managed
• session context, memory search/read, commit, and archive expansion behavior stays the same

The OpenViking service must be deployed and running independently before the plugin can connect to it.

Relationship to the Older Design Draft

The repo also contains a more future-looking design draft at docs/design/openclaw-context-engine-refactor.md. It is important not to conflate the two:

• this README describes current implemented behavior
• the older draft discusses a stronger future move into context-engine-owned lifecycle control
• in the current version, the main automatic recall path lives in assemble(): preflight rebuilds history, transformContext injects long-term memories
• in the current version, afterTurn() already appends to the OpenViking session, but commit remains threshold-triggered and asynchronous on that path
• in the current version, compact() already uses commit(wait=true), but it is still focused on synchronous commit plus readback rather than owning every orchestration concern

That distinction matters, otherwise the future design draft is easy to misread as already shipped behavior.

Operator and Debugging Surfaces

If you need to debug this plugin, start with these entry points.

Inspect the current setup

openclaw openviking status --json
openclaw plugins list
openclaw config get plugins.entries.openviking.config
openclaw config get plugins.slots.contextEngine

Watch logs

OpenClaw plugin logs:

openclaw logs --follow

OpenViking service logs:

cat ~/.openviking/data/log/openviking.log

Web Console

python -m openviking.console.bootstrap --host 0.0.0.0 --port 8020 --openviking-url http://127.0.0.1:1933

ov tui

ov tui

Common things to check

For installation, upgrade, and uninstall operations, use INSTALL.md.