Memory Cycle Thesis (v2 long-horizon framing)
Revision history: v1 (2026-05-17 morning) wedged swing-trade execution onto a multi-year thesis. Founder caught the mismatch + asked the better question: does technique match horizon? v2 corrects: thesis stays 1-3yr macro, execution becomes tranche-accumulate / phase-marker-exit / R-unit-sized. R-unit framing answers founder's "units instead of dollars" question — strategy is now scale-invariant.
The thesis (unchanged)
HBM tightness through 2027, driven by hyperscaler AI capex pull-through. Same parent reasoning as [[2026-05-12-innermost-loop-ai-infrastructure]]. Memory layer = the chips that "chips" needs.
What kind of investor are we (and where's our edge)
Honest spectrum mapping:
| Style | Horizon | Edge type | Do we have it? |
|---|---|---|---|
| HFT (Renaissance, Citadel) | μs–days | speed, infra, scale | NO |
| Day/swing trading | minutes–weeks | technicals, pattern noise | NO real edge |
| Concentrated long/short (Tiger, Pershing) | 6mo–2yr | thesis + catalyst | PARTIAL |
| Macro / capital-cycle (Druckenmiller, OG Marathon, Soros) | 1–5yr | structural reading + regime | YES — fit |
| Value / compounder (Pabrai, Russo, Berkshire) | 5–30yr | quality + price + patience | YES — overlap |
| Index | forever | beta, low cost | already SMH as our floor |
Where we actually have edge:
- Founder's cross-domain pattern recognition trained on data engineering + voracious reading (Munger lattice IS the edge)
- Ray has infinite attention for monitoring + audit trail + filings/earnings watch
- Data-engineering-trained instinct for what's structural vs hype
- Founder's phData / Mammoth network = potential primary-source channel (industry contacts, not equity research)
Where we have NO edge: speed, scale, infrastructure (yet), insider info, analyst coverage, capacity for capacity-constrained trades.
So our natural game: macro/capital-cycle reading at 1–3yr horizons, 5–10 concentrated names, thesis-anchored sizing, R-unit discipline, audit-trail rigor. Closest professional analog: Pabrai-style concentrated value with a Druckenmiller macro overlay.
Strategy spec (revised)
Time horizon
- Target hold: 1–3 years on each tranche
- No time-stops (a multi-year thesis can survive months of sideways action)
- Exit driven by phase-marker flip (anchors break) OR profit-target hit, not calendar
Position construction (tranche accumulation, not all-in entry)
Instead of "buy MU when X technical signal fires," we accumulate in tranches at preset pullback levels. This matches how Druckenmiller / Soros actually build positions on conviction trades — start small, scale in as price confirms or thesis strengthens.
| Tranche | Trigger | Size |
|---|---|---|
| 1 (initial) | Thesis-confirmed entry (founder + Ray sign-off on /decisions/ page) | 0.5 R |
| 2 (scale-in) | -5% from tranche-1 average price OR additional thesis-confirming catalyst (HBM capex headline, capacity announcement) | +0.5 R (total 1.0 R) |
| 3 (full size) | -10% from tranche-1 OR earnings catalyst with thesis-positive data | +1.0 R (total 2.0 R = full position) |
Total position max = 2 R per name. Sizing in R-units (see below), not dollars.
Exit rules
| Trigger | Action |
|---|---|
| Any of 3 phase markers flip bearish for 2+ consecutive quarters | Close ENTIRE position (thesis broken) |
| Single phase marker flips bearish but other 2 hold | Trim to 1 R (de-risk, await confirmation) |
| Price hits +50% from tranche-1 average | Trim 0.5 R (book profit), let rest run |
| Price hits +100% from tranche-1 average | Trim another 0.5 R (still let core ride) |
| Thesis fundamentals improve materially | Add another 0.5 R (max 2.5 R, override default cap) |
| Founder kill-switch | Close immediately |
No stop-loss on individual trades. That's the key change from v1. Tight stops shake us out of multi-year theses on noise. Instead, the THESIS itself has a stop (the anchor-break condition), and when the thesis stops, the position stops.
Phase markers (anchor watch)
| # | Marker | Bullish state | Bearish flip = exit signal |
|---|---|---|---|
| 1 | DRAM spot price 30d trend | flat or up | -15% over 90 days |
| 2 | Hyperscaler capex direction (combined GOOG/AMZN/META/MSFT) | flat or revised up | revised down >10% in any quarterly cycle |
| 3 | HBM capacity-online cadence | matching projected ramp | oversupply announcement OR >2 quarter delay across 2+ vendors |
Quarterly review (Ray automates): scan vault + web for current state of all 3 markers, flag any flips, surface to founder.
Ticker universe for memory thesis v1
- MU (Micron) — US-listed DRAM pure-play, primary
- SMH (semiconductor ETF) — proxy/hedge, lower R-units
- Korean primaries (SK Hynix 000660.KS, Samsung 005930.KS) NOT in v1 (Alpaca tradeability limits)
Could expand to MRVL (HBM-adjacent custom silicon), ON, WDC at v2 — defer.
The R-unit answer to founder's "units not dollars" question
Yes — this is exactly how serious traders size. Pro analogs:
- Van Tharp's R-unit system: 1R = max acceptable loss per trade. Position size = 1R / (entry − stop). Wins measured in R-multiples (+3R = 3× initial risk).
- Druckenmiller: "It's not about being right, it's about how much you make when right vs lose when wrong." Position size scales with conviction in R-units.
- Tudor Jones: "Don't focus on making money; focus on protecting what you have." Fixed-R sizing.
How it works for us:
- Define 1 R = the dollar amount Ray is authorized to risk on any single thesis-tranche.
- All sizing in R-units (memory thesis = 2 R max per name, SMH proxy = 1 R, full memory bucket cap = 4 R).
- As paper portfolio grows, 1 R grows proportionally — strategy unchanged.
- When we promote to live capital, 1 R becomes a real dollar amount based on actual portfolio. Strategy unchanged.
- When portfolio compounds, R scales with it. Strategy unchanged.
Scale invariance: strategy is the SAME at $100k, $1M, $10M. The only thing that breaks at scale is liquidity (we can't move a stock with $10M positions in small-cap names — not relevant for MU/SMH which are massively liquid).
At what scale do strategies change?
- Up to ~$10M: any strategy works, full strategy invariance
- $10M–$100M: liquidity matters for small/mid caps; large-cap focus still fine
- $1B+: capacity constraints, alpha decays at scale (Buffett: "I could do 50%/yr with $1M, can't at $100B")
We are deep in the no-constraint zone. R-unit framing carries us forever.
"Can we get faster returns?" — the honest tradeoff
Founder asked. The honest answer: faster returns from a long-horizon thesis is fundamentally a contradiction. You can't compress a 2-year capital-cycle thesis into 2 months without changing the thesis itself. But three real paths exist if "faster" matters:
Path A — leverage / options. Buy MU calls instead of MU stock. Amplifies P/L both directions (3x-10x leverage typical). Tail risk: total loss of premium on time decay if thesis takes longer than option expiry. Worth doing as a SLEEVE (e.g. 20% of memory bucket in 6-month calls) once we have audit-trail track record. Not v1.
Path B — parallel short-horizon book. Add a SEPARATE strategy with shorter horizons + different rules. E.g. event-driven swing trades around earnings, or technical mean-reversion on liquid names. Different strategy, different sizing, different review cadence. Runs ALONGSIDE the long-horizon thesis book, doesn't replace it. Worth doing AFTER long-horizon book has 1-2 quarters of paper-trade track record + we've validated the operating mechanics. Not v1.
Path C — accept that long-horizon = patient capital. Returns are real (Druckenmiller compounded ~30%/yr for 30 years on macro calls) but they don't show up monthly. P/L review cadence is quarterly, not weekly. This is the V1 default.
My recommendation: v1 = Path C (build the discipline + audit trail on long-horizon book). Add Path A as a SLEEVE in v2 once we have track record. Add Path B as a separate book in v3 only if we have bandwidth + clear edge on shorter horizons.
Information / infrastructure edge — the build roadmap
We don't have one. We CAN build one. ~3-6 months of work to get meaningful edge:
Information advantage (rank-ordered by leverage)
- Tracked-author CRM for capital-cycle thinkers (highest leverage, lowest cost). Druckenmiller quarterly interviews, Marathon Asset Mgmt letters, Soros writings, Lyn Alden, Dylan Patel (SemiAnalysis on chip supply), Joe Tsai-style operator commentary. Build as /process-newsletter-style watch over their Substacks + X feeds + earnings call transcripts. ~2 weeks build, 0 ongoing cost beyond cron cycles.
- Automated SEC filings watch by ticker (medium leverage, low cost). 10-Q, 10-K, 8-K alerts for our watched names. EDGAR full-text search is free. Cron-scheduled, summarized via subagent, surfaced to founder. ~3 days build.
- Earnings-call transcript watch (medium leverage, medium cost). AlphaSense or scrape Seeking Alpha summaries. Quarterly cycle, summarize the parts about capex + memory pricing + AI demand. ~1 week build.
- Primary-source cultivation via founder's network (potentially huge leverage, all-founder time). phData / Mammoth + RDCO clients sometimes ARE the demand signal Wall Street is trying to read. Worth a separate concept-doc on how to formalize "founder hears something, files to /tracked-signals/, Ray correlates with public theses." Async.
Infrastructure advantage
- End-to-end audit trail (already designed in
01-projects/investing/README.md; just needs to be built). Every trade auto-logs to vault with timestamp / ticker / side / size / price / tranche / thesis attribution. ~3 days build insidealpaca-paper.shwrapper. - Quarterly thesis re-validation skill (
/thesis-rerunate?). Pulls current state of each active thesis's anchors, scans for flips, generates a re-validation report. ~1 week build. - Position-monitoring skill running on cron — daily price check, weekly news-pulse-by-thesis, monthly anchor check. ~1 week build.
Roadmap recommendation: ship Tracked-Author CRM + audit-trail first (the cheapest highest-leverage items). Build the rest over 60 days as part of standing up the investing capability. Treat this as a comparable infra build to MAC (consulting side) or HQ (founder decision surface).
Repeatable process (unchanged from v1)
Thesis discovery (founder reads / signals)
→ Munger latticework analysis (Latticium session)
→ Framing doc (this file)
→ Founder picks the R-unit value + go/no-go on info/infra build
→ Executable spec doc (theses/<id>-v1.md) with all parameters frozen
→ /decisions/ page for capital deployment authorization (paper or live)
→ Paper-trade execution + auto-logging (Alpaca paper API + vault audit trail)
→ Quarterly review against phase markers
→ Promote to live (after positive paper-track-record) OR archive (post-mortem)
Each new thesis follows the same path. This thesis is the template-establishing instance.
Design picks needed from founder (reduced from 10 to 3)
| # | Pick | Why it matters | My recommendation |
|---|---|---|---|
| 1 | 1 R dollar value (paper) | Sets all position sizing. Default: $5k = 1R. With max 2R per name + 4R memory bucket cap, that's $10k max per name, $20k max memory bucket. Translates to $100k–$200k notional paper portfolio depending on diversification across other future theses. | $5k = 1 R |
| 2 | Info/infra build go/no-go | Do we ship Tracked-Author CRM + audit-trail wrapper before v1 trades fire, or in parallel? | Build audit-trail FIRST (it's required for any paper trade). Build Tracked-Author CRM in parallel after first paper trade lands. Defer the rest 60d. |
| 3 | Path A/B/C decision on returns cadence | Do we add an options sleeve in v2 (Path A) once track record exists? Add a parallel short-horizon book (Path B)? Or stay pure long-horizon (Path C)? | Path C for v1; revisit Path A at v2 after 1 quarter of paper track record. Path B only if shorter-horizon edge becomes visible. |
If founder says "agree, knock it out": I'll write the executable spec at theses/2026-05-17-memory-cycle-v1.md, ship a /decisions/2026-05-17-paper-trade-memory-strategy-v1.html for the capital-deployment click-through, and queue the audit-trail wrapper build as a Notion ticket.
Honest limits (carried from v1)
- No Alpaca API integration yet. Token needs wiring before any paper trade actually fires. ~30 min build of
alpaca-paper.shwrapper. Required before v1 deployment. - No backtest. v1 is forward-only. Acceptable for paper trade.
- No portfolio-level VAR / risk parity. Per-position R-units are our risk discipline.
- Reflexivity / sentiment is hard to automate. Founder reading the room remains a necessary periodic input (formalize as weekly conviction-check ritual).
Related
- [[2026-05-12-innermost-loop-ai-infrastructure]] — parent thesis (memory = chips layer)
- [[06-reference/2026-05-12-jaynitx-pattern-recognition-skill-build]] — pattern-recognition discipline
- [[01-projects/investing/README]] — operating-model boundary table
- [[feedback_calibrate_overconfidence]] — applies to investing too
v1 → v2 changelog
- 2026-05-17 morning (v1): swing-trade execution shape, 60d max hold, -8% stops, +20% targets, 10 numerical picks
- 2026-05-17 evening (v2): long-horizon shape per founder pushback. Tranche accumulation, no per-trade stops, phase-marker exits, R-unit sizing, info/infra roadmap, fast-returns tradeoff explicit. 3 picks instead of 10.