Files

m3tm3re 8bcb5e11fb feat(skills): add qmd skill for vault search and memory, deprecate SQLite memory

- 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.

2026-03-14 10:04:43 +01:00

5.0 KiB

Raw Blame History

name, description, compatibility

name	description	compatibility
qmd	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'.	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:

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.

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.

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.

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.

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:

~/CODEX/80-memory/{category}/{slug}.md

Required frontmatter:

---
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:

~/CODEX/80-memory/sessions/YYYY-MM-DD-{slug}.md

Frontmatter:

---
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.

Determine project/task context from working directory, git repo, and user goal.
Retrieve general relevant knowledge:

qmd query "{relevant context}" --md --full -n 3 -c CODEX

Retrieve project memory context:

qmd query "{project name}" --md -n 3 -c memory

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.

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

5.0 KiB Raw Blame History

qmd