Agent Memory Portability — 換 agent 要帶什麼

When to use

• Long debug / multi-hour coding; you will switch Claude ↔ Cursor ↔ Codex mid-task.
• Bigger context windows still fail: the new agent does not know what already failed.
• You want a file-first memory (carry bag), not a chat dump.

Problem framing

Context ≠ memory. A larger window still forgets decisions, failed attempts, and constraints when you change tools. Public signal: engineers ask what to “carry forward” when switching coding agents — not how to buy more tokens.

Memory model (three layers)

Prefer paths on disk over chat bubbles.

Carry-bag checklist (minimum export)

Copy into CARRY.md (or YAML) before switching:

1. Goal — one sentence
2. Constraints — languages, “do not touch”, security bounds
3. Tried & failed — bullet list with reasons
4. Key paths — files/dirs that matter now
5. Open questions — decisions still needed
6. Forbidden — commands/areas the next agent must not redo blindly

Minimal YAML shape

goal: "..."
constraints: ["..."]
failed:
  - attempt: "..."
    why: "..."
key_paths: ["src/...", "tests/..."]
open_questions: ["..."]
forbidden: ["force-push main", "rm -rf ..."]

Export / import protocol

1. End of session: write/update CARRY.md + optional DECISIONS.log (append only).
2. Point the next agent at those paths in the first message.
3. Optionally attach last failing test output path — not the whole transcript.
4. Harness tip: treat carry files as required reads at turn 0 and after every N turns.

Anti-patterns

• Pasting the entire chat history
• Relying only on a bigger context window
• Putting secrets / API keys into memory files
• “The model will remember” with no file

Acceptance test

After switching agents, one steering message that links CARRY.md is enough to continue — no full re-brief.

Output template

Goal:
Constraints:
Failed attempts:
Key paths:
Open questions:
Forbidden:
Next agent first prompt:

Out of scope

Building a full memory product; medical/therapy claims; publishing without operator approval.