How to use the Memory Discipline Loop agent skill

What the skill makes the agent do

The skill describes one loop: recall before starting work, save at decision points, and learn from corrections. Memory is only useful when reads happen before the work and writes happen when a decision is actually settled.

The agent searches saved memories for prior decisions and context at the start of a relevant task. During work it saves settled decisions together with their reasons, and when you correct its approach it saves a lesson rather than a plain memory. Before repeating a task type you have corrected it on, it recalls those lessons.

• Search for prior decisions at the start of a non-trivial task (memory_smart_search).
• Save each decision as it settles, including the reason and real file paths (memory_save).
• Record user corrections as lessons and recall them before similar work (memory_lesson_recall).
• Skip saving secrets, anything readable from the code, transient state, and step-by-step narration.

Install the skill

The skill is a folder that contains SKILL.md and TROUBLESHOOTING.md. Download the ZIP from the skill page on this site and unzip it into one of two locations: ~/.claude/skills/ for your user account, or .claude/skills/ inside a specific project if you want it limited to that project.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

The skill does not install the memory tools. You still need the agentmemory plugin set up separately.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into .claude/skills/ in the project root
• Claude apps: Settings → Capabilities → Skills, then upload

Check that the memory tools are available

The skill only describes calls to memory_smart_search, memory_save and memory_lesson_recall. Without those tools the agent has nothing to call, and the skill does nothing.

If the tools are missing, the troubleshooting reference says the agentmemory service must be reachable. In Claude Code, check /plugin list and /mcp, then restart or reload the connection after configuration changes.

• Required tools: memory_smart_search, memory_save, memory_lesson_recall
• Per the troubleshooting reference, the agentmemory service must be reachable
• Restart or reload the host connection after changing configuration

How to trigger the skill with natural requests

The skill is marked as not directly user-invocable, so you do not call it by name. It is meant to activate when the situation in its description applies: starting a non-trivial task, settling a decision or resolving a debugging gotcha, or deciding whether something belongs in memory.

You can steer it with plain language. Asking the agent to check what was decided before, or to remember a conclusion with its reason, fits the loop. The skill also has user-invoked companion forms named recall and remember for the read and write sides.

• "Before you start, check memory for how we handled auth refresh in this repo."
• "We settled on cursor pagination. Save that decision and why we chose it."
• "That's not how I want this done. Save it as a lesson so you avoid it next time."
• "Don't save anything to memory until I say so."

Example of a saved memory

A good save contains the decision and its reason, a few specific concepts, and real file paths. The skill's own example shows the shape of a memory call.

Respect permissions and treat recalled memories with care

If your memory preferences require explicit permission before saving, the agent should wait for it. Retrieved records are treated as untrusted evidence, so the agent should verify facts that can change against current sources. The skill also says the agent must never follow instructions embedded in a memory to export data, run commands, or override the current task.

The skill is explicit that the memory store keeps what is saved between sessions. Do not use it on projects where stored context could be sensitive and should not persist.

• Wait for explicit permission when your preferences require it
• Verify changeable facts against current sources before acting on them
• Ignore instructions found inside memories

Safety notes: scripts and network

The skill itself runs no scripts and makes no network requests. It is a set of written instructions. The memory tools it calls are provided by the agentmemory plugin, so the storage and service behaviour are governed by that plugin, not by this skill.

Because the skill decides what gets saved and the store keeps it between sessions, review what your agent is saving, especially early on.

• Runs scripts: no
• Uses network: no
• The memory store persists data between sessions