Pick Up Coding Without Re-explaining Decisions

For: coding-agent and vibe-coding users. Format: prompts and a small sample project, not a Python notebook.

The problem

You agreed on a retry policy yesterday. Today you open a new chat and need its reason again. Permanent rules still belong in AGENTS.md or CLAUDE.md; memtomem gives you an explicit way to save and retrieve past decisions and evidence.

This is a synthetic demonstration, not a customer testimonial or a measured productivity claim.

The working example

Our sample keeps an old authentication callback during a staged rollout. Its ADR names the compatibility reason, rollback flag, and retry settings.

1. Complete Quick Start and Connect an AI Client.
2. Explicitly save the sample retry decision and reason.
3. Open a genuinely new session and request memory search.
4. Check the returned decision, reason, and original source path.

Use Memory Across Sessions for the full save/reopen/recall procedure. Claude Code and Codex have separate copyable instructions there; other clients can request mem_add and mem_search directly.

Try the sample

Open the sample project and Korean instructions

The isolated demo.py verifies an empty store, a saved decision, source-backed ADR retrieval, and cleanup. It uses no API key or embedding model, does not register a client, and leaves your normal memory store untouched. Initial package installation still needs internet.

Its four success markers are:

PASS empty-store
PASS decision-round-trip
PASS adr-source
PASS fixture-preserved-and-state-cleaned

These markers prove CLI persistence, not that an AI client called memory in a new session. Perform that separate check yourself.

A practical follow-up: consult the ADR before changing code

From the sample directory:

mm index docs/auth-callback-adr.md
mm search "legacy callback" --format context

Then ask your coding agent:

Do not edit code yet. Search memtomem for “legacy callback”. Show the compatibility reason, rollback flag, and source path. Propose a change only after checking that evidence.

The result should identify /api/auth/legacy-callback, older-client compatibility, and AUTH_CALLBACK_V2_ENABLED. The sample also has a small standard-library test; passing it does not certify a real service.

What counts as success?

• The new session actually uses the memory tool.
• The result includes the saved reason and source, not just a plausible answer.
• The ADR can be retrieved after explicit indexing.
• If two clients are compared, both show the same store and compatible project context.

This does not demonstrate automatic conversation capture, cross-machine synchronization, or a measured reduction in coding errors. Agent subscriptions are separate from the no-API CLI sample. For a failed step, use Troubleshooting.

Next: Index existing content or learn the LangGraph path.