Memory Router
Decides where new content belongs in D Rage's JoVE knowledge system: Obsidian vault (RAM) for what changes, Pinecone (jove-memory index) for what doesn't. Apply this BEFORE writing any new artifact — not after.
The one routing rule
Ask: "Will this be edited again?"
- Yes → Obsidian vault (RAM)
- No → Pinecone
jove-memory(HD)
That's it. Everything below is tiebreakers and namespace-routing for when the answer isn't obvious.
Why this rule works
Obsidian and Pinecone are complementary, not competing — they serve different cognitive layers:
- Vault = reasoning surface. Mutable. Wiki-linked. Indexed by [[index]] + MOCs. The shape of D's thinking lives here. Every Claude session pays a token tax to load index.md + CLAUDE.md, so the vault should stay lean.
- Pinecone = perfect-recall archive. Immutable. Semantic search via E5 embeddings. Cost is flat regardless of size. Conversations, books, meetings, deep-research snapshots — everything that happened but isn't current belongs here.
Mixing them fails both ways: if you put archive content in the vault, the wiki bloats and the token tax grows linearly with content volume. If you put active strategy in Pinecone, you lose the wiki-link graph that makes reasoning possible.
Tiebreakers (apply in order when "Will this be edited again?" is ambiguous)
- Will it be searched semantically? ("Find anything we ever discussed about X.") → Pinecone. Vector search is the point.
- Will it be wiki-linked from MOCs or other notes? → Vault. The graph is the value.
- Is it a snapshot of state on a specific date? (e.g., "as of 2026-05-23, competitor X has Y features") → Pinecone with
as_ofmetadata. - Does it represent the current authoritative position? (e.g., "this is the PI persona we're going with") → Vault.
- Is it >2,000 words? → Strongly consider Pinecone chunking. Vault notes >2K words become unreadable.
- Does the user need to edit it by hand in Obsidian? → Vault.
- Is it generated by a deep-research / book-digest / meeting-transcript flow? → Pinecone (default for these flows).
Vault routing (when Obsidian wins)
Corrected 11 Aug 2026: the JoVE vault is ~/dev/jove-hq/, never ~/Obsidian/JoVE/. The old Johnny.Decimal vault was migrated into jove-hq and no longer exists on this machine; every routing below that named a NN — folder silently pointed at nothing. Layout authority is ~/dev/jove-hq/CLAUDE.md rule 12. Place per content type:
Location (under ~/dev/jove-hq/) |
Content |
|---|---|
knowledge/memory/ |
Durable working notes: who-is-who, standing preferences, registries, decisions that outlive a project |
knowledge/jove-labs/ |
Everything JoVE Labs: PRDs, decision memos, sweeps, the .sweep/findings.md ledger. INDEX.md there is generated, never hand-edited |
knowledge/meetings/YYYY-MM/ |
One coupled note per meeting; full transcripts → Pinecone meetings |
knowledge/entities/ |
People and org entities |
knowledge/copilot/ppp/ |
The PPP truth ledger and its drafts/<date>.md daily layer |
knowledge/copilot/ops/ |
Setup status, runbooks, manifests |
knowledge/copilot/wrap-ups/ |
Session wrap-up records (the /wrap-up skill writes here) |
product/, strategy/ |
Product-line and strategy artifacts |
source-docs/ |
Raw reference material, git-ignored for push |
Every vault note needs: YAML frontmatter per jove-hq convention (CLAUDE.md rule 5): type, status, links-in-frontmatter for entities, type and status tags only, no topic-tag sprawl. The old domain/confidence schema belonged to the migrated vault; do not reintroduce it.
Pinecone routing (when Pinecone wins)
Index: jove-memory. Embedding: multilingual-e5-large (integrated; 1024-dim; truncates at 512 tokens — keep chunks ≤350 words).
Pick the namespace:
| Namespace | What goes here | ID prefix pattern |
|---|---|---|
vault |
Do not write here. The auto-mirror of the jove-hq vault, owned by tools/vault-index/ingest.mjs with stable path#chunk ids; hand-upserts duplicate what the ingest owns. Verify with node tools/vault-index/health.mjs |
{repo-relative-path}#{n} |
conversations |
Wrap-ups of Claude sessions (≤350 words each) | convo-{YYYY-MM-DD}-{kebab-title} |
meetings |
Granola meeting transcripts + summaries (post-D-joins-JoVE) | meeting-{YYYY-MM-DD}-{kebab-title} |
emails |
Customer / Slack / institutional Gmail archives | email-{YYYY-MM-DD}-{kebab-title} |
research |
Deep-research outputs (competitive scans, market reports) | research-{topic-slug}-{YYYY-MM-DD}-{section} |
books |
Book digests (chunked by chapter/concept) | book-{title-slug}-{YYYY-MM-DD}-{section} |
daily-archive |
Daily notes >30 days old | daily-{YYYY-MM-DD} |
bootstrap-research |
FROZEN — 8 original research streams; do not write here | bootstrap-{stream}-{section} |
crawl-source |
FROZEN — JoVE website crawl; do not write here | crawl-{branch}-{key} |
Always include rich metadata per record: as_of, topic, tags[], source_urls[], plus any namespace-specific fields (e.g., branch for crawl, author for books, attendees for meetings).
Verify upserts: after upserting, do one mcp__pinecone__search-records round-trip on a query the record should match. Confirm hit at ≥0.80 similarity. If lower, the chunk is probably too noisy or too long.
The hybrid pattern (when both apply — the most common case)
For LARGE or IMMUTABLE content that also needs vault discoverability:
- Pinecone: full content, chunked appropriately. This is the canonical store.
- Vault: 1-page POINTER note that:
- Has standard frontmatter PLUS
pinecone_namespaceandpinecone_id_prefixfields - Lists the Pinecone record IDs as bullets
- Includes a teaser/synthesis paragraph (the "why this matters" condensed)
- Wiki-links to relevant MOCs, personas, features
- Cites 1 or 2 surprising findings ("the wedge insight")
- Has standard frontmatter PLUS
Existing examples to follow (the named notes below lived in the migrated vault; the pattern is what to copy, and current live examples of it are the corpus-digest pointer notes under knowledge/jove-labs/corpus-digests/):
- Hooked, Nir Eyal — 1-page book pointer with 8 Pinecone IDs in
booksnamespace - Crawl Provenance — pointer to 287 records in
crawl-source - Competitors, AI Assessment for Science — competitor pointer with 6 records in
research - Competitors, Lab Hiring Funnels — same pattern
The pointer note is what shows up in the vault graph + MOC indexes; the Pinecone records are the answer-substrate for semantic queries.
Decision examples (apply the rule)
Example 1: User says "Ingest the 'Nature Protocols' deep-dive I just researched."
→ Will it be edited again? No (it's a snapshot of state on 2026-05-23).
→ Pinecone research namespace + vault pointer note Competitors — Nature Protocols.md. Hybrid.
Example 2: User says "Update the PI Persona note to add a 6th sub-persona." → Will it be edited again? Yes (personas evolve as PI conversations happen). → Vault edit only. No Pinecone.
Example 3: User says "Save this conversation about the architecture decision we just made."
→ Will it be edited again? No (conversation wraps are by definition done).
→ Pinecone conversations namespace via /wrap-up skill. Log entry to vault.
Example 4: User says "Create a Granola meeting note for today's call with the CEO."
→ Will it be edited again? The summary stub MAYBE (action items get checked off); transcript NO.
→ Hybrid: coupled note in knowledge/meetings/2026-MM/2026-MM-DD-ceo-1on1.md (action items + decisions + key quotes ≤200 words); full transcript in Pinecone meetings.
Example 5: User says "Help me draft a new strategic decision about JoVE Labs pricing."
→ Will it be edited again? Yes (decisions get refined; review_date is in frontmatter for a reason).
→ Vault decision note: knowledge/jove-labs/DECISION-MEMO-<topic>-<date>.md for Labs, knowledge/memory/ for cross-project calls. No manual Pinecone; the mirror ingests it.
Example 6: User says "Index Alex Hormozi's $100M Leads book."
→ Will it be edited again? No.
→ Pinecone books namespace + vault pointer stub. Hybrid.
Example 7: User says "Add this article I just read to the vault."
→ Apply tiebreaker 6: will user edit by hand? Probably not (it's a clipped source).
→ Pinecone research namespace + vault pointer if the article matters to a live topic. Otherwise source-docs/ (raw reference, kept out of git pushes per jove-hq CLAUDE.md rule 8).
Anti-patterns (do NOT do these)
- ❌ Don't put curated active strategy (MOCs, decisions, personas, product anchors) in Pinecone. Vault IS RAM; the point is mutation + wiki-link graph.
- ❌ Don't put full meeting transcripts in vault. They bloat the wiki. Stub-in-vault + transcript-in-Pinecone is the pattern.
- ❌ Don't create both a vault note AND a Pinecone record where the vault note IS the canonical state — that creates 2 sources of truth that drift. Use the hybrid pattern only when content is large + immutable AND needs vault discoverability.
- ❌ Don't backfill existing vault content into Pinecone. Vault is RAM; backfill would duplicate, not enrich.
- ❌ Don't write API keys, secrets, or credentials to either layer.
- ❌ Don't put a
as_ofsnapshot into the vault as if it were current. Tag itas_ofand route to Pinecone.
Output format
When applying this skill, state the routing decision EXPLICITLY before executing — one line:
Routing decision: <vault path> [+ Pinecone <namespace> <id-prefix>]. Why: <one-sentence reason citing the rule or tiebreaker>.
Then execute (write the file / upsert the records).
Source of truth (read these for deeper context)
~/dev/jove-hq/ARCHITECTURE.md§ Memory tiering — the live namespace map and routing design~/dev/jove-hq/CLAUDE.mdrule 4 — the enforced two-layer rule: vault mutable, Pinecone = auto-mirror (vaultnamespace) plus immutable haystack~/.claude/skills/wrap-up/SKILL.md— companion skill for theconversationsnamespace flow
When NOT to use this skill
Skip the explicit routing decision for:
- Single-question Q&A with no artifact produced
- Trivial vault edits (typo fixes, frontmatter date bumps)
- Reading existing content (search/recall path, not write path)
- Pure conversation that hasn't produced anything to store