Case study: Slide deck vs GitHub — memory_recall, cache, and filing parity issues (April 2026)

This case study walks through a Cursor session that aligned the ClawQL pitch deck in docs/clawql-slides.md with the GitHub backlog. The goal was to close narrative gaps by creating tracking issues for themes that appeared in the 80-slide deck (including §08 security) but had no or incomplete issue coverage.

Audience: Teams using ClawQL MCP who want a repeatable split between durable context (memory_ingest / memory_recall / Obsidian), ephemeral session state (cache, optional, in-process), and GitHub for execution.

---

1. Why this session mattered

A consolidated slide deck is a product contract: it lists Web3, Fabric, The Graph, Chainlink, OSV-Scanner, Istio, defense-in-depth, and more. If that narrative runs ahead of issues and code, the story drifts. This session was reconcile deck → open issues → new issues while keeping a clear boundary:

Layer	Tooling	Role in this session
Ephemeral scratch	MCP cache (when CLAWQL_ENABLE_CACHE=1)	Stash transient keys (working notes, last-known issue list, “do not put in the vault” scratch) — not long-term memory.
Durable recall	memory_recall (vault Markdown)	Before writing new issues, pull prior decisions and deck-related notes so filing is informed, not from chat memory alone.
Durable write	memory_ingest	After the session, append a summary of what was created and how it links to the deck.
Execution	gh, GitHub	The system of record for shippable work.

For when cache is vs is not the right tool, see docs/cache-tool.md. cache is in-process and LRU-bounded — it is the right place for “temporary issue numbers and information we do not want to persist to long-term memory” for as long as the MCP server process (and the session) lasts.

---

2. The slide deck as product contract

• Source: docs/clawql-slides.md — 80 slides, sections §01–§08 (Core through Defense in Depth), plus a closing slide.
• Paired long-form security reference: docs/clawql-security-defense-in-depth.md (aligns with §08).
• Not used as a second source of truth: the deck is aspirational in places; the gap pass explicitly preferred “file an issue and label scope” over silently assuming the repo already matched every slide bullet.

The assistant enumerated all open issues in danielsmithdevelopment/ClawQL, then compared that set to the deck’s major themes (Istio, OSV, Fabric, The Graph, Chainlink, Jaeger, Vault, Web3, transcript parity, etc.).

---

3. memory_recall before creating issues

Before opening new GitHub issues, the assistant called memory_recall (ClawQL MCP) with a focused query — e.g. keywords like clawql slide deck, slides security, roadmap — so the filing pass could reuse prior vault notes instead of inventing context.

What came back (examples):

• Memory/clawql-slide-deck-and-security-documentation-april-2026.md — prior ingest describing the 80-slide deck and §08 work.
• Memory/clawql-complete-consolidated-slide-deck-april-2026.md and transcript-related notes (older; useful as history, not as the canonical slide count).
• Memory/clawql-agent-platform-vision-and-roadmap-2026-04-17.md and related roadmap material.

Why this step matters: memory_recall grounds “what we already said in the vault” (titles, wikilinks, earlier prioritization) so new issues can link to stable narrative and avoid duplicate “deck refresh” work where an ingest already exists. It does not replace gh issue list, which is the authoritative set of open engineering tickets.

Parameters used (illustrative): query = short keyword phrase, limit small (e.g. 5–8), optional maxDepth when graph follow-through matters (wikilinks). See docs/mcp-tools.md under memory_recall.

---

4. cache for short-lived working state

During the gap pass, the assistant used the MCP cache tool to hold transient state that should not be committed to the Obsidian vault — for example:

• A short working summary of “what we are comparing this round” (deck section → theme list).
• Scratch keys such as deck-gap-analysis-2026-04-25 and, after new issues were filed, a compact list of new issue numbers and titles under something like deck-gaps-issues — ephemeral labels for the rest of the same assistant run, not durable project memory.

Important: cache is ephemeral in this MCP server process (see docs/cache-tool.md). A restart, a different process, or another client does not see these keys. That is the intended behavior: scratch, not a substitute for memory_ingest.

Contrast: Issue numbers, titles, and the rationale for each gap were written to GitHub (durable, linkable) and, at the end, to memory_ingest (durable, recallable in the vault). The cache copy was only to avoid re-listing 12 issues in every tool turn within the session.

---

5. Comparing the deck to open issues

1. Extract structure from the deck: slide titles / sections (e.g. Architecture, Web3, Defense in Depth) and major themes.
2. List open issues in the repo (gh issue list --state open — sort by number for a stable view).
3. Keyword map: for each major theme (Istio, OSV, Fabric, Graph, Chainlink, x402, …), check whether an open issue already tracks it. Many were already covered (e.g. public x402 / gateway #88, Helm full-stack was previously shipped; Onyx / Flink have dedicated closed or open items).
4. Gap = deck claims work that has no (or a clearly incomplete) issue — file a new issue with scope, acceptance, and links to the deck and related docs so future readers know why the issue exists.

---

6. New GitHub issues filed (April 2026)

Twelve new issues were opened to connect the April 2026 deck to an explicit backlog (numbers shown as filed in that session):

#	Area
#155	Optional Istio + Kiali, mTLS east-west
#156	OSV-Scanner + Golden Image / CI / optional CronJob / search+execute path
#157	Hyperledger Fabric — Helm, CLAWQL_ENABLE_FABRIC, providers/fabric
#158	The Graph — bundled OpenAPI / execute path
#159	Chainlink — bundled provider surface
#160	Jaeger / OTLP tracing
#161	HashiCorp Vault or OpenBao (vs chart Obsidian vault hostPath naming)
#162	ClawQL-Web3 — AgentKit / IPFS / CCIP (extends beyond #88)
#163	Transcript parity: clawql-slides-transcript.md vs 80 slides + §08
#164	Defense-in-depth doc → control / deliverable matrix
#165	Meta: update obsolete slide § references in existing issue bodies
#166	Demos — honest walkthroughs for high-stakes narrative slides (e.g. 50, 56)

Cross-links between issues (e.g. #88, #132, #133, #128, #129, #131) were added in the bodies so P2/P3 dependencies stay navigable from GitHub, not only from the deck.

---

7. memory_ingest when the work is done

After creating issues and updating the mental model of “deck ↔ backlog,” the assistant used memory_ingest with a stable title (so related sessions append), a sessionId, and append: true, to record:

• The 12 new issues and the thematic buckets.
• A pointer to the canonical deck file and the defense-in-depth companion doc.
• (Optional) wikilinks to related vault note titles, e.g. the consolidated deck ingest note, for graph recall in Obsidian.

Why not only cache? The vault is the durable place to answer, next month: “What did we file when the deck went to 80 slides?” cache from that day is long gone. memory_ingest is the durable counterpart for human+assistant continuity.

---

8. Scratch, trail, and durable memory (summary)

Store	When to use	This session
cache	Session handoff, scratch keys, ephemeral lists (same MCP process)	Stashed working gap labels and new-issue list in-session; not vault-persisted.
audit	In-process operator trail of MCP tool calls (optional)	Not the focus; see docs/enterprise-mcp-tools.md and #89.
memory_recall	Find existing deck/roadmap/security notes before editing or filing	Queried the vault for slide-deck and prior security ingests before creating issues.
GitHub issues	Shippable, linkable, assignable work	12 new issues + cross-links to prior epics.
memory_ingest	Durable summary of outcomes and wikilinks for the graph	Appended session outcome to the long-running “slide deck + security” note (or a sibling title).

---

9. References

• docs/clawql-slides.md — canonical deck
• docs/clawql-security-defense-in-depth.md — §08 long-form
• docs/cache-tool.md — cache tool semantics
• docs/mcp-tools.md — memory_ingest, memory_recall
• GitHub: issues #155–#166 in danielsmithdevelopment/ClawQL (April 2026)
• .cursor/skills/clawql-vault-memory/SKILL.md — deep ingest and recall pattern

Readable layout (this site): docs.clawql.com/case-studies/slide-deck-github-parity-cache-memory-recall-2026-04 — keep in sync with the repo docs/case_studies/ file above.

---

Case study metadata

Item	Value
Session	April 2026, Cursor + ClawQL MCP (scoped server id e.g. project-0-ClawQL-clawql in multi-config setups)
Outcomes	12 new GitHub issues, vault memory_ingest append, this case study + website mirror
Primary lesson	memory_recall for grounding → gh for ground truth on open work → cache only for same-session scratch that must not bloat the vault → memory_ingest for durable “what we decided and filed”