One brain. Seven agents. Zero cold starts.
A single Notion knowledge base that every LLM and agent reads from and writes back to. Each session starts with the full accumulated context instead of an empty prompt, so judgement compounds across every model rather than resetting each time.
Architecture
What we chose and why. The source of truth for every meaningful buy, sell, build, config or process call.
What we learned. Research, signals, lessons and source material that back the decisions.
How agents behave. Reusable operating rules every agent reads before acting.
What ran and what it produced. Automation health and output trace across the fleet.
It compounds
The memory is almost free to carry
The corpus keeps growing as every agent writes back. The context an agent loads to get oriented stays small and flat, because agents read a compact operating contract and then query specific decisions and entries on demand instead of swallowing the whole base.
Cost to carry
- DeepSeek V4 Flash· primaryload$0.00041k$0.41cached$0.008 / 1k
- Gemini 3.5 Flashload$0.00441k$4.37cached$0.44 / 1k
- Claude Sonnet 4.6load$0.00871k$8.74cached$0.87 / 1k
Input only, o200k tokenizer on the live operating pages. Input list prices verified late May 2026. Prompt caching cuts repeat loads by roughly 90 percent.
The four layer model
Each layer has exactly one job
The record of what was chosen. Every meaningful action lands here with a rationale and an outcome.
The evidence behind the calls. Graded by reliability so a filing never carries the same weight as a screenshot.
The rules agents load before acting. Tax aware investing rules, automation cadence, risk caps, handoff rules.
The weekly control room. Action queue, decision radar, stale thesis review, evidence inbox, automation health.
One source of truth per concept. No concept lives in two layers at once.
How it grew
Emergent, not designed
There was no master plan. It started as one page. Each stage only appeared when the previous one stopped scaling. The inspiration was Karpathy's LLM wiki pattern, pushed from one agent to seven.
- One pageA place to capture thoughts. No schema, no plan.
- Multiple pagesMore topics, more pages. It started to sprawl.
- First databasePages could not scale. Structure earned its place.
- Entries schemaEach unit of knowledge became a real record.
- Tagscompounding starts hereRecords connected across domains and became reusable.
- System ConfigRepeated lessons hardened into rules agents load first.
- Seven agent meshEvery model reads and writes the same base.
- Frontier restructureGPT 5.5 and Opus 4.8 periodically rebuild it.
The differentiator
Most people query their knowledge base. I use a frontier model to rebuild mine.
Every few weeks, once enough has piled up, GPT 5.5 and Opus 4.8 read the whole base. Not to answer a question, to reorganise it. The model proposes schema changes, merges and new links. I debate it, approve a plan, and implement only what earns its place. The architecture was negotiated, not designed.
The moat is not the model. It is the memory and the control layer you build around it.
The agent contract
Ten rules every agent follows
What stops seven different models from drifting into seven different versions of the truth.
- Search first. read existing context before creating anything new.
- Ground in Config. read relevant System Config before giving advice.
- No duplicates. update canonical pages instead of spawning near copies.
- Link, do not orphan. tie new work to existing decisions and entries.
- Supersede cleanly. mark old views superseded, never silently overwrite.
- Investing rigour. include tax impact, invalidating signal, confidence, review date.
- Flag stale. if source freshness is uncertain, say so directly.
- No raw to decision. never promote a loose note without a clear rationale.
- No secrets. no passwords, keys, tokens or credentials ever stored.
- Close the loop. every handoff carries owner, output, evidence, failure handling.
Weekly consolidation
Pruning, not just collecting
- Collectsweep the week of outputs across all seven agents.
- Dedupeupdate canonical pages first, no copy spam.
- Promoterepeated lessons graduate into System Config.
- Prunearchive stale entries, mark superseded views.
- Nextend with one action doable in under 30 minutes.
Closing
Built and maintained by Vishal Shah.