# Jarvis Memory > Use when a user asks Jarvis to remember, forget, correct, or consult durable context, or when a Jarvis workflow surfaces a stable candidate. - Skill: `erionis/jarvis-memory` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add erionis/jarvis-memory` - Raw SKILL.md: https://api.skillmd.com/api/skills/erionis/jarvis-memory/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: Erionis (https://skillmd.com/u/erionis) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/erionis/jarvis-memory --- # Jarvis Memory Act as the semantic curator of `Identity` and `Durable memory`, not a global gateway. Bootstrap, briefing, and ordinary work may consult declared sources without invoking this workflow. Direct reads stay direct. Consultation is read-only and never authorizes a change. Resolve `Durable memory` through the consumer's declared capability map and resolve `Identity` the same way. Paths and names can differ. If a role required for the current intent is missing, ambiguous, or points to a missing source, stop without inventing a fallback; block only that action. ## Classify before proposing Search the existing content, index, and likely primary source first. Keep each fact in one authoritative source—one authoritative home: - behavior that should apply across every session and domain -> `Identity`; - domains, repositories, tools, assets, or rituals -> stable local context; - stable decisions, preferences, ownership, constraints, and traps -> `Durable memory`; - knowledge specific to one body of work -> project-owned knowledge; - unfinished actions -> `Future work`; - completed events -> `Daily history`; - versions, availability, and current progress -> live state or its living source; - credentials, tokens, and keys -> an existing secret store or a protected local file. Curate only `Identity` and `Durable memory` directly. Route every other category to the workflow that owns its authoritative source. The governing boundary is **do not copy live state** into durable memory. Jarvis Doctor remains read-only; its findings become separate curation proposals, never automatic repairs. ## Preview every persistent change An explicit request authorizes a proposal, not a write. Before every change to `Identity` or `Durable memory`, show: 1. the semantic destination, with the path only when useful or requested; 2. the proposed text or readable diff; corrections use before/after; 3. why it belongs there, duplicates avoided, and links affected; 4. the canonical runtime choice UI with `Save`, `Edit`, or `Do not save`, or numbered equivalents when no choice UI exists. There is **no write before confirmation**. Silence, refusal, or an edit request leaves the files unchanged. For deletion, full overwrite, or deep semantic restructuring, also apply the guardrail's impact, recovery, and double confirmation requirements. Always **propose before semantic deletion**: name the exact fact and its authoritative location, explain impact and recovery, and obtain explicit approval before modifying it. An inferred pattern is only a candidate. During ordinary work, offer at most one high-value stable candidate at a natural boundary. When another workflow delegates signals, evaluate at most five candidates in one compact proposal. A pending or rejected proposal does not block the calling workflow. The caller retains ownership of ordering, checkpoint status, the final response, and Git. Git stays outside this skill. Do not stage, commit, push, or change Git configuration. ## Grow structure progressively Keep the bootstrap-loaded memory short and update facts in place. When recurrence and durable density form a coherent topic with no better home, propose an indexed detail note with its name, content to move, and index patch; never create it automatically. Recurrence is evidence, not a numeric threshold. If a project hub exists, the project hub remains authoritative. Durable memory keeps only orientation and a pointer. Moving a general principle into Identity is a semantic move, not a copy: keep unique triggers and exceptions in the subordinate detail. Raw uploads stay in `Inbox` until their destination is clear or confirmed; preserve original files. When work reveals a customer or project, propose an existing or approved domain location and its hub; the owning workflow performs the move. Do not impose a universal `Projects/raw/deliverables` hierarchy. ## Apply an approved proposal safely 1. Re-read the target immediately before the patch. 2. If it changed, preserve non-overlapping changes. On semantic overlap, or if it changes again, reject last-writer-wins and do not write. 3. Apply the smallest approved patch to the current version. Update indexes or pointers only when the preview included them. 4. Re-read the result and verify every local link touched. 5. Report the semantic target, result, and verification without reopening the decision. ## Secrets and informed override Warn by default and do not retain a secret in Identity or Durable memory; **do not repeat the value**. **Do not display the value in previews, final responses, or diagnostic commands**; never inspect a secret-bearing target with a command or diff that prints its contents. Prefer the consumer's existing secret store or a local destination verified as excluded from Git, then hand the approved action to that destination's owning workflow. If the user insists on a named tracked target, explain persistence in Git history and future remotes, show the exact path without the value, and require a second explicit confirmation equivalent to `save anyway` in a **separate user turn after the warning**. The initial request cannot satisfy this second gate, even when it already anticipates the risk or says `save anyway`. After that informed confirmation, comply through the owning workflow. Before asking for it, state honestly that **runtime or tool history may retain the supplied value** because the write operation must carry it; do not promise technical-log redaction that the runtime cannot guarantee. Keep every avoidable display and all user-facing output redacted.