kdb/LLM Memory System.md

# LLM Memory System

## Purpose

This vault maintains **persistent knowledge** that helps the LLM understand:

1. **Who you are** — Your expertise, goals, working style
2. **How you want to collaborate** — Feedback, preferences, constraints
3. **Project context** — Ongoing work, decisions, deadlines
4. **External resources** — Tools, links, references

## Why This Matters

Without persistent memory, each conversation starts from scratch. This system allows the LLM to:
- Remember your preferences across sessions
- Build on previous decisions
- Avoid repeating mistakes
- Provide context-aware assistance

## How It Works

### File-Based Storage

All memories are stored as **Markdown files** with YAML frontmatter:

```yaml
---
name: example_memory
description: Short description
type: feedback
created: 2026-05-15
---
```

### Bi-directional Linking

Memories link to related notes using `[[wikilinks]]`:

```markdown
See also: [[tier2-setup-retro]], [[bumakopania-research]]
```

### Tagging System

Use tags for filtering and querying:
- `#type/user`
- `#type/feedback`
- `#type/project`
- `#type/reference`

## Memory Types

### User Memories (`memories/user/`)

About **you** — your role, expertise, goals, preferences.

**When to create:**
- Learning about your background
- Understanding your goals
- Noting your expertise level

**Example:**
```markdown
---
name: senior-kdb-developer
description: Senior developer working with KDB/time-series data
type: user
---
```

### Feedback Memories (`memories/feedback/`)

**Guidance** on how the LLM should work with you.

**When to create:**
- After corrections ("no, not that")
- After confirmations ("yes, exactly")
- When establishing preferences

**Structure:**
- **The Rule** — What to do/avoid
- **Why** — Reason behind it
- **How to Apply** — When it matters

**Example:**
```markdown
---
name: no-summary-responses
description: Keep responses terse, no trailing summaries
type: feedback
---

## The Rule
Don't summarize what you just did at the end of every response.

## Why
User can read the diff; summaries are redundant.

## How to Apply
After making code changes, just show the diff. No "I've updated X, Y, Z" summary.
```

### Project Memories (`memories/project/`)

**Ongoing work** — decisions, context, deadlines.

**When to create:**
- Starting new initiatives
- Recording decisions
- Noting deadlines/constraints

**Structure:**
- **Fact/Decision** — What's happening
- **Why** — Motivation
- **How to Apply** — Impact on work
- **Timeline** — Dates, status

**Example:**
```markdown
---
name: tier2-setup-2026-05-14
description: Setting up local knowledge database (kiwix-serve + zim-llm)
type: project
status: active
---
```

### Reference Memories (`memories/reference/`)

**External resources** — tools, links, documentation.

**When to create:**
- Finding useful resources
- Recording tool configurations
- Noting where to find information

**Example:**
```markdown
---
name: offline-docs-sources
description: Sources for offline documentation (Kiwix, LibGen, etc.)
type: reference
url: https://download.kiwix.org/
---
```

## Workflow

### Adding New Memories

1. **Identify type** — user, feedback, project, or reference
2. **Create note** — Use appropriate template
3. **Fill frontmatter** — Name, description, type
4. **Add content** — Rule/fact, why, how to apply
5. **Link related** — Add `[[wikilinks]]` to connected notes

### Finding Memories

1. **Graph view** — Visualize connections
2. **Search** — Ctrl/Cmd + Shift + F
3. **Tag pane** — Filter by type
4. **Backlinks** — See what links here
5. **Smart Connections** (plugin) — Semantic search

### Maintaining Quality

- **Be specific** — Clear rules, concrete examples
- **Include why** — Context helps with edge cases
- **Update regularly** — Remove outdated info
- **Link generously** — Build knowledge graph

## Integration with LLM

### Automatic Loading

The LLM loads `MEMORY.md` and all `.md` files in this directory at the start of each session.

### Manual Recall

Ask the LLM to:
- "Check memory for X"
- "What do you know about Y?"
- "Remember that Z"

### LLM Plugins

For advanced features:
- **Smart Connections** — Vector search
- **ObsidianLLM** — AI-assisted writing
- **Templater** — Dynamic templates

## Best Practices

1. **Save immediately** — Don't wait to add feedback
2. **Be concise** — One idea per note
3. **Use templates** — Consistent structure
4. **Link everything** — Build the knowledge graph
5. **Review regularly** — Archive outdated memories

## Migration from Previous System

This system builds on the existing file-based structure:
- Existing `.md` files remain in root
- New memories go in `memories/{type}/`
- Templates ensure consistency
- Obsidian adds linking, search, visualization

---

**Related:** [[MEMORY.md]], [[README.md]]