Use when the user explicitly invokes Pensyve, asks what the agent remembers, wants Codex to recall or store project memory, requests memory health review, or wants memory grounded work.
Installs into .claude/skills of the current project.
Are you the author of Pensyve?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/major7apps-pensyve)
---
name: pensyve
description: Use when the user explicitly invokes Pensyve, asks what the agent remembers, wants Codex to recall or store project memory, requests memory health review, or wants memory grounded work.
---
# Pensyve
Use Pensyve as Codex's persistent working memory. The skill is intended for explicit `$pensyve` invocation from Codex, selection through `/skills`, `/pensyve` command workflows, text-level `@pensyve` compatibility requests, and implicit routing when a task clearly depends on prior project memory.
## When To Use
- The user asks what you remember about a project, repo, decision, person, or workflow.
- The user asks you to remember a fact, preference, constraint, root cause, or reusable procedure.
- You are about to make a substantive recommendation that prior project decisions could affect.
- The user asks to review, clean up, or inspect memory state.
- The user asks Codex to continue prior work without re-briefing.
- The user types `@pensyve` as a mention-style memory request.
Do not use this skill for trivial commands, simple formatting, one-off shell output, or facts that should not be persisted.
## Workflow
1. Detect the memory entity.
- Prefer the repo, product, service, file, module, or user-named subject.
- Use lowercase hyphenated entity names.
- If the entity is ambiguous, ask one short clarifying question or fall back to the repository name.
2. Recall before substantive claims.
- Use `pensyve_recall` with a query that includes the task and entity.
- Keep recall output concise: summarize relevant memories rather than dumping raw records.
- If no memory is found, say that briefly only when it matters.
3. Store only high-signal memory.
- Use `pensyve_remember` for durable facts, decisions, preferences, and constraints.
- Use `pensyve_observe` for session outcomes, root causes, failed approaches, and procedural lessons.
- Prefix procedural observations with `[procedural]`.
- Strip API keys, tokens, passwords, private URLs, and credentials before writing.
4. Review or inspect when asked.
- Use `pensyve_inspect` for entity-specific memory listings.
- Use `pensyve_status` for connection or namespace health.
- Ask for confirmation before using `pensyve_forget`.
## Error Handling
- If Pensyve MCP tools are unavailable, say: "Pensyve MCP is not connected for this session." Then continue without memory.
- If a write fails, report the failed item and continue with any remaining requested work.
- Never fabricate memories. Only present what Pensyve returned or what the current session established.
- Do not claim true `@pensyve` platform dispatch. Current Codex support is `$pensyve`, `/skills`, `/pensyve`, and text-level `@pensyve` interpretation by the model.