memory-recall

Memory Recall

Surface what the user already knows from their vault before answering from training data. The vault is the source of truth for personal decisions, accumulated learnings, and prior context.

Recall is index-first (Karpathy's LLM-wiki principle). Read the catalog (index.md) first to find relevant pages, then drill into those pages. qmd is a search accelerator used only when the index alone is not enough. This is faithful to the wiki pattern and works better than blind semantic retrieval at this scale.

If $ARGUMENTS is provided (e.g. /memory-recall what do I know about useEffect), that string is the topic. Otherwise infer the topic from the conversation.

Tool roles (the hybrid)

• index.md = the curated catalog. Read FIRST. The navigation map: titles, one-line summaries, tags, [[wikilinks]].
• qmd = content search across the whole vault. Use as a FALLBACK when the index skim is ambiguous or to catch pages whose body matches but whose summary does not.
• obsidian CLI = navigation. Follow [[backlinks]], resolve aliases, read a known page, list a folder, read the live vault (no index lag).

Recall Flow (Mode A — topic recall)

Step 1 — Read the index first (always)

Read the catalog by ABSOLUTE path so this works from any directory, not just the vault:

Read /Users/kriscard/obsidian-vault-kriscard/index.md

Skim the ## Concept notes (claude-memory) section (and ## Articles / ## Videos if relevant). Pick 3–8 candidate pages whose title, one-line summary, or tags match the topic. This selection is reasoning over the catalog, not keyword scoring.

Step 2 — qmd only if the index is not enough

If the index skim is ambiguous, returns nothing obvious, or the topic might live in page bodies not reflected in summaries, search content:

qmd query "<topic>" --json -n 8 2>/dev/null

The single-line form auto-expands (qmd generates lex/vec/hyde internally). Escalate for harder cases:

• Ambiguous term — this vault mixes domains, so "performance" could be React render perf, Hyrox fitness, or team health. Add intent to steer expansion and reranking toward the sense you mean.
• Unknown vocabulary — add a hyde: line: 50–100 words of the answer you expect.
• Max recall — structured lex: + vec:; put the best guess first (it gets 2x fusion weight).

qmd query $'intent: react render performance, not fitness\nlex: useEffect rerender\nvec: how to avoid unnecessary renders' --json -n 8 2>/dev/null

Add the qmd hits to the candidate set. Score bands: ≥ 0.7 strong, 0.5–0.7 probable, < 0.5 weak.

qmd gotchas: (1) progress goes to stderr, so always append 2>/dev/null to keep --json output clean for parsing. (2) In lex/vec/hyde lines a hyphen inside a word like re-render is parsed as -negation and errors; drop the hyphen or rephrase.

Step 3 — Drill into the actual pages (navigation = obsidian CLI)

Read the candidate pages in full before answering. Use obsidian CLI to read the live page and follow the wiki's link structure; use qmd get "#docid" for a quick body pull of a qmd hit.

obsidian read path="3 - Resources/<subfolder>/<page>.md"
qmd get "#<docid>"

Follow [[backlinks]] and resolve aliases via obsidian to pull in connected pages the catalog/search missed.

Step 4 — Synthesize with citations

Answer from the pages, citing each with [[wikilink]]. Synthesize across pages; do not dump full notes unless asked. Be honest about gaps: if no index entry matches AND qmd top score < 0.5, say the vault has no coverage and offer a general-knowledge answer.

Step 5 — Offer to file the answer back

If the synthesis is worth keeping (a new connection, comparison, or decision), offer to file it as a wiki page via the save-note skill. Good answers compound into the wiki.

Connection Discovery Mode (Mode B)

When the user asks to connect two domains ("connect A and B", "bridges between"):

1. Map each domain index-first: read index.md, collect the pages under each domain by title/tag; augment with qmd query "<domain>" --json 2>/dev/null if the catalog is thin.
2. Pull hubs and follow backlinks 2–3 hops via obsidian CLI. If one domain has far fewer notes, go 3–4 hops deep on the sparse side — that is where surprises are.
3. Find overlaps: shared references (notes in both backlink chains), shared themes (same concept in both, even if unlinked), shared patterns (both stuck on the same problem). A useful qmd trick: feed one domain's vocabulary as a hyde: query against the other to surface adjacent pages.
4. Synthesize each bridge:

Bridge [#]: [Title]
  In Domain A: [how it appears]
  In Domain B: [how it appears differently]
  The connection: [what links them]
  Depth: Surface / Structural / Foundational
  Implication: [what this suggests for either domain]

Surface the strongest bridge and any missing links (connections that should exist but have not been made). The test: the best output makes you see both domains differently. Do not force connections; if the domains genuinely do not connect, say so.

How to Report Findings

• Cite the source with the file path / [[wikilink]] so the user can navigate.
• Synthesize, do not dump. Summarize what the pages collectively say.
• Be honest about gaps. No index match and qmd < 0.5 means say so, then offer general knowledge.
• Distinguish note types: TIL (3 - Resources/TIL/) = first-person captures; Resources = wiki reference; Projects (1 - Projects/) = project-specific context.

What NOT to do

• Don't skip the index. Index-first is the rule; qmd is the fallback, not the entry point.
• Don't write or modify notes — recall is strictly read-only (filing back is the save-note skill).
• Don't answer from snippets alone — read the pages that matter first.
• Don't hallucinate citations — if you can't find a note, say "no match found".
• Don't stretch a sub-0.5 match into a confident answer — flag the gap instead.