Skip to main content

openviking-memory

Recall and persist long-term memory through the OpenViking MCP tools. Use at the start of any substantive task (coding, configuration, debugging, multi-step or tool-based work) to retrieve relevant prior knowledge with find/search/read, and during or after work to persist durable facts, preferences, decisions, and lessons with remember. Do not use for casual chat or simple factual questions the model can answer directly.

Informations de source

Dépôt
volcengine/OpenViking
Dernière activité de la source
1 octobre 2026 à 07:00
Langue détectée de SKILL.md
anglais
Étoiles
39 649
Forks
3 132

Options d'installation

Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.

Vérifiez les fichiers source

Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.

Explorateur de fichiers
2 fichiers

Affichage de SKILL.md

SKILL.md
Instructions source · Aperçu en lecture seule
name
openviking-memory
description
Recall and persist long-term memory through the OpenViking MCP tools. Use at the start of any substantive task (coding, configuration, debugging, multi-step or tool-based work) to retrieve relevant prior knowledge with find/search/read, and during or after work to persist durable facts, preferences, decisions, and lessons with remember. Do not use for casual chat or simple factual questions the model can answer directly.
# OpenViking Memory OpenViking is a long-term semantic memory store addressed by `viking://` URIs. This client has no lifecycle hooks, so nothing is recalled or captured automatically — you drive both halves of the loop with the `openviking` MCP tools. Core tools, available on every supported deployment: - Recall: `find`, `search`, `read`, `list`, `grep`, `glob` - Persist: `remember`, `add_resource` - Maintain: `forget`, `health` Some deployments register more than the core set — `tree`, `write`, `edit`, `list_watches`, `cancel_watch`. These are optional: which ones exist depends on the server version and hosting mode (the managed cloud service trims some of them). Check the session's registered tool list; if any optional tool is present, read [references/optional-tools.md](references/optional-tools.md) before using it. Never call a tool that is not registered, and do not fall back to raw HTTP. If no OpenViking tools are registered at all, continue without memory. ## Recall: at task start 1. Decide whether the request warrants memory. Retrieve for executable or multi-step work, anything touching a system you may have seen before, and recovery from failures. Skip retrieval for small talk and one-off trivia. 2. Build one concise query from the task goal, domain objects, intended operation, and constraints. After a failure, include the failed operation and the stable part of the error message. 3. Call `find` (fast, ranked results with URI + abstract + score) with `limit` around 5-10. Use `search` when deeper intent analysis helps, or use `search` with `mode="context"` for a server-assembled, token-budgeted context block. In list mode, scope with `target_uri` when you know where to look, e.g. `viking://~/memories/experiences` for prior task experience. `viking://~` is the home alias for your own user root; a server that predates the alias rejects every `viking://~` URI with `INVALID_URI`. Against such a server use the explicit `viking://user/<user_id>/...` root taken from a URI already visible in this session, or drop `target_uri` and keep the hits whose URI contains `/memories/experiences/`. Never guess a user ID. 4. Judge results by task and environment fit, not title similarity. `read` the one to three exact file URIs likely to change how you execute. Ignore sidecar files such as `.abstract.md`, `.overview.md`, and `.relations.json`. 5. If nothing relevant comes back, proceed without memory. Make at most one focused follow-up search when execution fails for a materially new reason. Treat retrieved memory as advisory. Priority order: system and developer instructions, the current user request, current environment and tool evidence, then memory. Verify commands, paths, and versions against the present task; prior success never authorizes a destructive action now. ## Persist: during and after work Because capture is not automatic here, durable information is lost unless you store it. When you encounter something worth keeping, persist it in the same session: - `remember(messages)` — the default. Pass the key exchange or a short factual summary as role-tagged messages; the server extracts and files memories (preferences, entities, events, experience) on its own. Use it when the user says "remember this", states a lasting preference or decision, or when a hard-won lesson (root cause, working procedure, environment quirk) emerges. - `add_resource` — to import external documents or URLs as searchable resources. Use it for any file, URL, or repo the user hands you; never copy its text into a file with `write` instead. - When you need a note you author at a known location (under `viking://~/` — your own user root — or `viking://resources/`), the optional `write` / `edit` tools cover that — see [references/optional-tools.md](references/optional-tools.md). If they are not registered, fall back to `remember`. What to persist: stable preferences and conventions, environment facts, decisions with their rationale, and reusable procedures or fixes. What not to persist: secrets and credentials, transient state, speculation, or bulk transcript dumps — store conclusions, not scrollback. ## Example User asks to fix a failing deployment: 1. `find` with query `deployment image pull failure private registry`, `target_uri: "viking://~/memories/experiences"`. 2. `read` the most relevant experience URI; check its assumptions against the current cluster before applying its steps. 3. Fix the issue, verify the live result. 4. `remember` a short summary of the root cause and the working fix so the next session can recall it.
Voir sur GitHub