remember

SKILL.md

AI has no memory between sessions. Every new session starts blank. This skill fixes that.

Run it at the end of a session to save. Run it at the start of a new session to restore. That is all it does — but done consistently, it means nothing ever gets lost.

Security Boundary

This skill must never persist secrets. If any sensitive value appears in the conversation or context, do not copy it to memory.md.

Sensitive data includes (non-exhaustive):

API keys, access tokens, refresh tokens, session tokens

Passwords, passphrases, one-time codes, private keys, certificates

Cookies, auth headers, connection strings, webhook secrets

Any credential-like value or secret-looking string

If a detail is useful but sensitive, store a redacted placeholder instead (for example: [REDACTED_API_KEY]).
If unsure whether something is sensitive, treat it as sensitive and omit or redact it.

How to Invoke
To save at end of session:

/remember save

To restore at start of new session:

/remember restore

If the developer just runs /remember without specifying — ask them which one they need.

Save Mode

When the developer runs /remember save:

What to capture

Review the current conversation to extract only what a developer would genuinely need to continue this work in a completely fresh context. Do not include sensitive data such as credentials, API keys, or tokens in the saved memory. Not a transcript. Not a summary of everything that happened. The essential state.

Think like someone handing off a project to a colleague who is equally skilled but knows nothing about what happened today. What would they need to know to continue without losing anything?

Capture:

What was built — specific files created or modified, features completed, components added. Be precise. Not "built the auth flow" — "created app/(auth)/login/page.tsx, app/(auth)/callback/page.tsx, and middleware.ts. OAuth with Google and GitHub working end to end."

Decisions made — choices that would be hard to reverse or that future work depends on. Not implementation details — architectural choices. "Chose to use server-side data fetching over client-side — avoids loading states and keeps sensitive logic off the client."

Problems solved — any issue that took time to figure out. So the next session does not solve the same problem twice. "Third party auth callback requires a trailing slash in the redirect URL — fixed in the callback handler."

Current state — exactly where things stand right now. What works, what is partial, what is known to be broken.

What comes next — the very next thing that needs to happen. Specific enough that the next session can start immediately without figuring out where to begin.

Open questions — anything unresolved that the next session needs to address.

What not to capture

Implementation details that are visible in the code

Decisions already documented in context files

Anything that can be inferred by reading the codebase

The process of how something was built — only what was built and what was decided

Any secrets or credential-like values (tokens, keys, passwords, cookies, auth headers, connection strings)

Safety check before writing

Before writing memory.md, run a final pass over the content to ensure no sensitive value is present.

If sensitive content is found, remove or redact it before writing.

Keep only the minimal non-sensitive context needed to continue next session.

Where to save

Write the memory to memory.md in the project root. This file always contains only the most recent session state.

If memory.md already exists, show the developer a brief summary of what is currently saved and ask for confirmation before overwriting:

Step 1 — Read memory.md, provide the one-line summary, and stop to wait for developer input:

memory.md already exists from a previous session.
Current memory covers: [one-line summary of existing content].

Overwrite with this session's memory? (yes / no)

Step 2 — After the developer responds:

If they say yes, write the new memory.md.

If they say no, do not write anything and reply:

No changes made. memory.md is unchanged.

Format

# Memory — [Feature or Session Name]

Last updated: [date and time]

## What was built

[Specific files, components, features completed this session]

## Decisions made

[Architectural and implementation decisions that future work depends on]

## Problems solved

[Issues resolved this session — so they are not solved again]

## Current state

[Exactly where things stand — what works, what is partial, what is broken]

## Next session starts with

[The very first thing to do in the next session — specific and actionable]

## Open questions

[Anything unresolved that needs addressing]

After writing the file, confirm to the developer:

Memory saved to memory.md.

Next session: run /remember restore to pick up from here.

Restore Mode

When the developer runs /remember restore at the start of a new session:

Step 1 — Find the memory

Look for memory.md in the project root. If it does not exist, tell the developer:

No memory.md found in this project.

Either this is the first session, or the file was not saved.
To save memory at the end of a session, run /remember save.

Step 2 — Read everything available

Read memory.md first. Then check for these specific context files if they exist and read only those:

CLAUDE.md, .claude/context.md — Claude Code

.github/copilot-instructions.md — GitHub Copilot

.cursorrules, .cursor/rules/ — Cursor

.windsurfrules — Windsurf

AGENTS.md — Codex

.clinerules — Cline

context.md — generic fallback

Do not scan or read other files beyond this list. Build the most complete picture possible from what is available.

When restoring, never repeat or surface raw secrets from any source. If a secret appears in restored context, summarise it in redacted form only.

Step 3 — Confirm what was restored

Do not start building. Do not assume the developer wants to continue immediately. Summarise what was restored so the developer can verify Claude understood correctly.

Memory restored. Here is where we are:

**Last session:** [what was built]
**Current state:** [what works right now]
**Decisions in place:** [key decisions that are locked]
**Next up:** [what the next session should start with]

Is this correct? Say yes to continue, or correct anything
that does not look right before we proceed.

Only after the developer confirms does the session continue.

If memory is incomplete or unclear

If memory.md exists but is missing important context, say so honestly:

I found memory.md but some context seems missing —
[what is unclear or absent].

Should we continue with what we have, or do you want
to fill in the gaps before we start?

Do not guess. Do not assume. Surface the gap and let the developer decide.

The Rule

Every session ends with /remember save.
Every session starts with /remember restore.

That is the whole system. Consistent use is what makes it work.
A skill used sometimes is a skill that cannot be relied on.