save-learnings

name	save-learnings
description	Propose and persist approved Agents Memory updates for the current project. Use when the user invokes /save-learnings or asks to save project learnings, decisions, architecture changes, instructions, or obsolete memories from the current conversation.

Save Learnings

Persist only user-approved project memories. Never auto-save.

Workflow

1. Load existing project memory: ~/.agents-memory/bin/agents-memory get --cwd "$PWD" --all --format json
2. Review the current conversation and existing entries.
3. Propose numbered operations using only:
   • create
   • update (revise an existing active entry in place)
   • archive (drop an active entry from startup injection; kept for reference)
4. Each create or update must be atomic and include:
   • type: instruction, decision, or observation
   • priority: high, medium, or low
   • content
   • rationale
5. Ask for explicit approval. Accept partial approval such as approve 1,3,4.
6. Infer the current agent as codex, claude, or opencode.
7. Apply only approved operations by passing JSON to: ~/.agents-memory/bin/agents-memory apply --cwd "$PWD" --agent <agent> --operations-file -
8. Report the tool result.

Proposal Format

Use this shape:

Proposed memory changes:

[1] create decision, high
Content:
  ...
Rationale:
  ...

[2] update #4, medium
Before:
  ...
After:
  ...
Rationale:
  ...

[3] archive #7

Reply with `approve all`, `approve 1,3`, or `reject`.

Apply JSON

Send a JSON object with an operations array:

{
  "operations": [
    {
      "action": "create",
      "type": "decision",
      "priority": "high",
      "content": "...",
      "rationale": "..."
    },
    {
      "action": "update",
      "target_id": 4,
      "type": "decision",
      "priority": "medium",
      "content": "...",
      "rationale": "..."
    },
    {
      "action": "archive",
      "target_id": 7
    }
  ]
}

Rules

• Always respond in the same language the user is writing in.
• Do not store transcripts or conversation summaries.
• Use update to revise an active entry in place; use archive to drop one from startup.
• Do not persist unapproved operations.
• Do not add files to the project.
• Do not use unknown as the agent if the current agent can be inferred from runtime context.
• If the internal tool is missing, tell the user to run /setup-agents-memory.