memory-automation

Files (1)

• SKILL.md

SKILL.md (first lines)

# Memory Automation

Automates the persistent-memory workflow for Hermes using Mnemon graph memory. Used by `.hermes.md` instructions to keep every session primed with context.

## Mnemon Quick Reference

### Mnemon Tool Signatures

```python
mnemon_remember(text, category, importance, entities, tags)  # Plugin tool — direct call
mnemon_recall(query, intent, limit)                          # Plugin tool — direct call
mnemon_forget(insight_id)                                    # Plugin tool — direct call
```

These are real Hermes plugin tools, not pseudocode. See [Save Patterns](#save-patterns) for when to use them directly vs via delegated subagent.

### Categories (for `category` parameter)

| Category | When to use |
|----------|-------------|
| `fact` | Objective information about the user, project, or environment |
| `preference` | User's likes, dislikes, stylistic choices, workflow preferences |
| `decision` | Architectural decisions, tool choices, design decisions |
| `insight` | Non-obvious findings, troubleshooting steps, workarounds |
| `context` | Session-level context that's useful across sessions |
| `general` | Anything that doesn't fit above |

### Importance Levels

| Level | When to use |
|-------|-------------|
| 5 | Critical — user identity, core preferences, security constraints |
| 4 | Important — project goals, recurring preferences, key constraints |
| 3 | Normal — useful context, typical insights |
| 2 | Minor — nice-to-know details |
| 1 | Trivial — barely worth saving (use rarely) |

## Recall Patterns

### Full Context Load (Session Start)

```python
# Called by the SessionStart hook
mnemon_recall(query="", intent="GENERAL", limit=20)
```

This loads the 20 most relevant past insights into context. No specific query needed — just a broad sweep of everything mnemon has stored.

### Topic-Specific Recall (Before Responding)

```python
# If the hook returns nothing but the topic references past work:
mnemon_recall(query="<topic keywords>", intent="GENERAL", limit=10)
```

Use the user's message topic as the query. Include key terms, project names, and any domain-specific vocabulary.

### Intent-Targeted Recall

```python
# When you need to understand WHY a decision was made:
mnemon_recall(query="<topic>", intent="WHY", limit=5)

# When you need to know WHEN something happened:
mnemon_recall(query="<topic>", intent="WHEN", limit=5)

# When you need info about a specific entity:
mnemon_recall(query="<entity name>", intent="ENTITY", limit=5)
```

## Save Patterns

### Direct Tool Call (Preferred for Simple Saves)

`mnemon_remember` is a real Hermes plugin tool exposed by the `gitricko/hermes-plugin-mnemon` plugin. Call it directly from your toolset — no subagent, no CLI invocation needed:

```python
mnemon_remember(
    text="User prefers concise responses and bullet points over prose.",
    category="preference",
    importance=4,
    entities=["gitricko"],
    tags=["communication-style","user-preference"]
)
```

**Use direct calls when:**
- Saving a single fact or preference inline during a response
- The save is straightforward — no extraction, no multi-item batch
- You're already in a turn; just batch the tool call with whatever else you're doing

The tool wraps the `mnemon` CLI. No MCP server, no subagent.

### Delegated Save (Bulk or After Complex Responses)

For bulk saves (3+ items) or when context is tight, delegate to a subagent with `toolsets=["terminal"]`. The subagent uses the `mnemon remember` CLI since plugin tools may not propagate to child sessions:

```python
delegate_task(
    goal="Save insights to mnemon memory",
    context="""Save the following to mnemon memory:

1. text="User prefers bullet-point summaries over paragraphs.",
   category="preference", importance=4,
   entities=["gitricko"], tags=["communication-style"]

2. text="The project uses FastAPI with PostgreSQL.",
   category="fact", importance=3,
   entities=["project-name"], tags=["tech-stack"]

For each item run: mnemon remember "<text>" \\
  --cat <category> --imp <N> \\
  --entities e1,e2 \\
  --tags "t1,t2"
Binary: /usr/local/bin/mnemon.""",
    toolsets=["terminal"]
)
```

From gitricko/.minions, MIT.

More from gitricko/.minions