m3tam3re/AGENTS
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(skills): add qmd skill for vault search and memory, deprecate SQLite memory

Browse Source 
- Add skills/qmd/SKILL.md: qmd-based knowledge retrieval and memory system
  - Search workflows: BM25, vector semantic, hybrid+rerank modes
  - Memory write patterns via direct filesystem writes to ~/CODEX/80-memory/
  - Session summary workflow for capturing work at session end
  - Auto-recall pattern for context injection at session start

- Deprecate skills/memory/SKILL.md: SQLite-based memory replaced by qmd
  - qmd provides superior hybrid search over entire vault
  - No external OpenAI embedding dependency
  - Uses local GPU (7900 XTX) for embeddings and reranking

- Update skills/obsidian/SKILL.md: remove memory search sections
  - Memory search now handled by qmd skill
  - Keep CRUD operations and daily notes functionality
  - Update integration table to reference qmd

Related: CODEX vault updated with memory template, session-summary template,
and sessions directory. qmd collections and context annotations configured.
This commit is contained in:
m3tm3re
2026-03-14 10:03:53 +01:00
parent 39ac89f388
commit 8bcb5e11fb
2 changed files with 199 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
skills/obsidian/SKILL.md
View File

@@ -218,7 +218,8 @@ curl -X POST "http://127.0.0.1:27124/create-note" \
| Research note | research | Save research findings with tags | | Research note | research | Save research findings with tags |
| Project note | task-management | Link tasks to project notes | | Project note | task-management | Link tasks to project notes |
| Plan document | plan-writing | Save generated plan to vault | | Plan document | plan-writing | Save generated plan to vault |
| Memory note | memory | Create/read memory notes in 80-memory/ | | Vault search | qmd | Search vault content via qmd hybrid search |
| Memory note | qmd | Create/read memory notes in 80-memory/ |
## Best Practices ## Best Practices
@@ -232,97 +233,15 @@ curl -X POST "http://127.0.0.1:27124/create-note" \
--- ---
## Memory Folder Conventions ## Memory System
The `80-memory/` folder stores dual-layer memories synced with Mem0. Memory storage and retrieval is handled by the **qmd skill** (`skills/qmd/SKILL.md`).
### Structure - **Search**: Use qmd for searching vault content (hybrid BM25 + vector + reranking)
- **Write**: Use this skill's REST API or direct filesystem writes for creating/updating notes
- **Structure**: `80-memory/` folder with subdirectories (preferences, facts, decisions, entities, other, sessions)
``` See the qmd skill for memory workflows, session summaries, and auto-recall patterns.
80-memory/
├── preferences/ # Personal preferences (UI, workflow, communication)
├── facts/ # Objective information (role, tech stack, constraints)
├── decisions/ # Choices with rationale (tool selections, architecture)
├── entities/ # People, organizations, systems, concepts
└── other/ # Everything else
```
### Naming Convention
Memory notes use kebab-case: `prefers-dark-mode.md`, `uses-typescript.md`
### Required Frontmatter
```yaml
---
type: memory
category: # preference | fact | decision | entity | other
mem0_id: # Mem0 memory ID (e.g., "mem_abc123")
source: explicit # explicit | auto-capture
importance: # critical | high | medium | low
created: 2026-02-12
updated: 2026-02-12
tags:
- memory
sync_targets: []
---
```
### Key Fields
| Field | Purpose |
|-------|---------|
| `mem0_id` | Links to Mem0 entry for semantic search |
| `category` | Determines subfolder and classification |
| `source` | How memory was captured (explicit request vs auto) |
| `importance` | Priority for recall ranking |
---
## Memory Note Workflows
### Create Memory Note
When creating a memory note in the vault:
```bash
# Using REST API
curl -X POST "http://127.0.0.1:27124/create-note" \
-H "Content-Type: application/json" \
-d '{
"path": "80-memory/preferences/prefers-dark-mode.md",
"content": "---\ntype: memory\ncategory: preference\nmem0_id: mem_abc123\nsource: explicit\nimportance: medium\ncreated: 2026-02-12\nupdated: 2026-02-12\ntags:\n - memory\nsync_targets: []\n---\n\n# Prefers Dark Mode\n\n## Content\n\nUser prefers dark mode in all applications.\n\n## Context\n\nStated during UI preferences discussion on 2026-02-12.\n\n## Related\n\n- [[UI Settings]]\n"
}'
```
### Read Memory Note
Read by path with URL encoding:
```bash
curl -X GET "http://127.0.0.1:27124/read-note?path=80-memory%2Fpreferences%2Fprefers-dark-mode.md"
```
### Search Memories
Search within memory folder:
```bash
curl -X GET "http://127.0.0.1:27124/search?q=dark%20mode&path=80-memory"
```
### Update Memory Note
Update content and frontmatter:
```bash
curl -X PUT "http://127.0.0.1:27124/update-note" \
-H "Content-Type: application/json" \
-d '{
"path": "80-memory/preferences/prefers-dark-mode.md",
"content": "# Updated content..."
}'
```
--- ---

191
skills/qmd/SKILL.md Normal file
View File

@@ -0,0 +1,191 @@
---
name: qmd
description: "Knowledge retrieval and memory system via qmd (Query Markup Documents). Use when: (1) Searching knowledge vault for relevant context, (2) Storing memories, decisions, preferences from sessions, (3) Auto-recall of relevant context at session start, (4) Writing session summaries, (5) Querying past decisions or project context. Triggers: 'remember', 'recall', 'search vault', 'what did we decide', 'session summary', 'qmd', 'knowledge', 'memory', 'context'."
compatibility: opencode
---
# qmd
Primary retrieval + memory workflow using qmd v2.0.1 over the already-indexed `qmd://CODEX/` vault.
## Prerequisites
- `qmd` installed globally and available in `PATH`
- `CODEX` collection configured and indexed
- Vault at `~/CODEX/` with memory folders under `80-memory/`
Verify runtime state:
```bash
qmd status
```
## Search Workflows
Default to `qmd query` unless a narrower mode is clearly better.
### 1) BM25 Keyword Search (Fast)
Use when exact terms, filenames, tags, or known phrases are likely present.
```bash
qmd search "query"
qmd search "query" --md --full -n 5
qmd search "query" --json -n 10
qmd search "query" -c memory --md -n 5
qmd search "query" --min-score 0.3 --all --files
```
### 2) Vector Semantic Search
Use when the concept is known but wording is uncertain.
```bash
qmd vsearch "query"
qmd vsearch "query" --md --full -n 5
qmd vsearch "query" --json -n 10
qmd vsearch "query" -c memory --md -n 5
qmd vsearch "query" --min-score 0.3 --all --files
```
### 3) Hybrid + Rerank (Best Quality, Default)
Use for high-confidence retrieval, ambiguous queries, and session bootstrap.
```bash
qmd query "query"
qmd query "query" --md --full -n 5
qmd query "query" --json -n 10
qmd query "query" -c memory --md -n 5
qmd query "query" --min-score 0.3 --all --files
```
## Document Retrieval
Use retrieval commands after search to pull exact documents.
```bash
qmd get "path/to/file.md"
qmd get "#docid"
qmd multi-get "80-memory/sessions/*.md"
qmd multi-get "file1.md, file2.md"
```
## Memory Write Workflows
qmd is read-only. Write memory notes directly to filesystem, then re-index.
Memory note file pattern:
```bash
~/CODEX/80-memory/{category}/{slug}.md
```
Required frontmatter:
```yaml
---
type: memory
category: preference # preference | fact | decision | entity | other
source: opencode # opencode | manual
importance: medium # critical | high | medium | low
session: # opencode session ID for traceability (optional)
project: # [[Project Name]] if project-specific (optional)
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags:
- memory
- preference
---
```
Content template:
- `# {Title}`
- `## Content` → concise factual memory
- `## Context` → when/why captured
- `## Related` → WikiLinks to related notes
Category decision guide:
| Signal | Category | Example |
|--------|----------|---------|
| "I prefer...", "I like...", "Always use..." | preference | Prefers Nix over Docker |
| "We use...", "The stack is...", "My role is..." | fact | Tech stack is NixOS + TypeScript |
| "Let's go with...", "We decided...", reasoning about choices | decision | Chose qmd for memory retrieval |
| Person name, org, system, API | entity | Tobi Lütke - Shopify CEO |
| Doesn't fit above | other | |
## Session Summary Workflow
Write at session end or when explicitly requested.
Session file pattern:
```bash
~/CODEX/80-memory/sessions/YYYY-MM-DD-{slug}.md
```
Frontmatter:
```yaml
---
type: session-summary
created: YYYY-MM-DD
session: # opencode session ID
project: # [[Project Name]]
duration: # approximate
tags:
- memory
- session
---
```
Content template:
- `# Session: {Brief Description}`
- `## What Was Done` → completed work bullets
- `## Decisions Made` → decision + short rationale (+ WikiLink if significant)
- `## Open Questions` → unresolved items
- `## Files Changed` → key file list + one-line deltas
- `## Next Steps` → next executable actions
## Auto-Recall Workflow
At session start, proactively retrieve context before deep work.
1. Determine project/task context from working directory, git repo, and user goal.
2. Retrieve general relevant knowledge:
```bash
qmd query "{relevant context}" --md --full -n 3 -c CODEX
```
3. Retrieve project memory context:
```bash
qmd query "{project name}" --md -n 3 -c memory
```
4. Inject only relevant excerpts; cap recall to ~2000 tokens.
This workflow is advisory and should be run proactively by the agent.
## Re-indexing
After creating or editing memory/session files, update index for immediate recall.
```bash
qmd update
```
Indexing also runs periodically; use manual update when immediate searchability is required.
## Integration with Other Skills
| From | To | Pattern |
|------|----|---------|
| qmd search | Any skill | Retrieve relevant context before starting work |
| brainstorming | qmd write | Save brainstorm decisions as memory notes |
| Any session | qmd write | Capture session summaries at session end |
| qmd search | obsidian | Find note to update, then use Obsidian skill for CRUD |