The brain is LIVE (2026-06-12): ghostteam-brain repo, cloned to each founder’s home-directory top level. It is the system this page described as a need — we built our own rather than adopting G-Brain/Paperclip wholesale.
Operating model: agent.md instructs every agent to git pull --rebase on session start and push after writes; harness-agnostic (agent.md + claude.md are all any agent needs — Codex, Claude Code, Hermes). Wiki is one page per customer / company / person, plus pipeline and product pages. Granola transcripts feed an ingest-call skill: the agent proposes wiki updates, a human reviews and commits — no blind agent writes while trust is being built (expected load ~2–3 calls/day, one review each — manageable).
The Alpic→“lpic” misspelling carried from a test-run transcript into the wiki is the canonical proof that the human review gate is needed early.
Product wiki is the hardest and highest-priority page: complex, context-heavy, no single right answer (active web, auth, feature scope), and most relevant context never appears in transcripts — so it needs deliberate ongoing human curation. Customer pages are easier (agent ~90% right, humans spot-check). The GitHub codebase itself is useful raw context for the product page.
Brain-folder integrity is a live risk: opening the brain in a coding session could flood it with sales docs / work artifacts. Safeguard = agent.md rules that block non-brain content + a scratch/ folder that temp files redirect to — but discipline is still required alongside the rules.
Skills are centralized in the brain, not duplicated per person (e.g. one “create Ghost Team deck” skill, used by both founders).
Agent/harness strategy is deliberately not fixed yet — 1–2 weeks to get a feel before scaling. Current split: Vincent leans Claude for writing/planning, Elliot uses Codex for coding/automation. A 24/7 Hermes agent for proactive/automated call ingestion is the next phase once the brain is stable; Slack (over Telegram) becomes relevant once multiple 24/7 agents run. Slack-vs-Telegram decision deferred.
Defined need (2026-06-05): a shared company memory system with permissions, usable from ANY client/harness (Codex, Claude Code, Hermes, etc.) — “consistent and aggregated view of everything we’re doing”, not locked to one coding agent’s own memory.
Architecture both founders liked: git-based memory with a human review gate (updates land as PRs that the founders approve); future employees/agents get scoped access. Legal/sensitive sections may need an access gate — unsolved.
Options surveyed: “G-Brain” (open-source memory file system attributed to Garry Tan / YC; GitHub-based with PR review gates — possibly optimized for OpenClaw/Hermes and small teams) and Paperclip (open-source agent orchestration; very active — 95 contributors, updated within the hour; agent profiles/permissions, backlog-like planning UI). Vincent’s counter-idea: build on Google Drive for its mature permissioning.
Agreed approach: start small, test one feature rather than adopting a whole orchestration platform; avoid hyped tools — prefer projects still actively maintained months after the hype.
Driver for 24/7 access: Codex requires the computer to stay on for mobile access — impractical for Elliot.
Hermes supports multiple personalized profiles on one VPS — possible basis for CTO/CMO-style specialized agents; practical question open: how to invoke a profile inside another harness (e.g. @-ing an MCP per prompt risks bloating context; memory pull should happen once at session start).
Idea-vs-fact convention already exists in the research subfolder (clarified 2026-06-26). The brain already has a research/experiment subfolder with status-tagged tables (e.g. signal-set-and-weights, which Elliot calls “the primitive version of the recommendation algorithm”) — Elliot built it; the morning’s “where is this captured?” friction was founder awareness/sync, not a missing structure. Agreed convention: ideas enter the table tagged as ideas; promotion to fact requires explicit evidence (an agent reads the status column). Corollary to the rule below — both founders must know what already exists and update each other, or they re-derive structure in parallel.
Critical-infra facts must not be silently defined by transcript ingestion (flagged 2026-06-26). Elliot queried the brain for “the four R’s and the factors we measure” and got inconsistent answers across phrasings; he suspected a stale/wrong auto-ingested “ADO methodology” page. No such page exists — the canonical four-R definition lives in discoverability-framework with bucket mechanics in ado-score-model, so the issue was retrieval/context, not a bad page. Rule going forward: the platform + canonical product pages are the single source of truth for the score’s factors + weightings; full transcripts shouldn’t define critical infra. Action: Vincent to locate-or-recreate the ADO methodology presentation (~3 weeks old, lost across Telegram/Cursor) as one clean canonical overview.
Open questions
Best way to inject brain context into active working sessions (Codex, Claude projects) without polluting the folder — options floated: Claude projects folder, global instructions, MCP/CLI reference. Elliot to share a proposed approach for review, then align.
Permission/authorization model for sectioned memory (legal locked to founders) — design needed.
“Airbit” mentioned as another option in the company-context/memory space — confirm name and evaluate.
(Resolved 2026-06-12) “G-Brain”/Paperclip survey closed — we shipped our own ghostteam-brain rather than adopting either.
How should ingest distinguish truth vs idea vs hypothesis so the wiki stays trustworthy as “truth” (ideas/hypotheses captured but clearly tagged — e.g. the signal-set candidate table), and how do we keep critical definitions (four R’s, score factors) from drifting via transcript ingestion?
Timeline
[2026-07-01] (internal discussion — Vincent + Elliot afternoon call) Brain-structure alignment that seeded the later epistemic standard (2026-07-06-epistemic-status-standard): agreed the wiki holds three foundational documents — (1) research areas / hypotheses to test, (2) the “recommendation brain” (factors that drive discovery + rank), (3) signals & weights — plus standalone insight docs that accrue evidence over time (first candidate: the retrieval-rank ↔ picker correlation). Vincent’s sprint task stated verbatim: propose a way to track ongoing research differentiating ideas, evidence, and facts, and create a research backlog from the Q2 plan. Also agreed [idea]: an editable web layer on top of the GitHub wiki (wiki.) because important docs can’t be “vibe-committed” and GitHub isn’t live-typeable — Elliot wants click-and-type granularity. Scope re-affirmed: the brain is for the founders + their agents to get smarter — not something the platform is built on; chosen over Google Docs/Telegram/Canva precisely because it’s permanent and agent-readable (“agents just read it — that’s exactly what we want”).
[2026-06-26] (internal deck-review — Vincent + Elliot, recorded by Elliot) Resolved part of the morning brain-trust gap: the research/experiment subfolder + idea-vs-fact status tables already exist (Elliot built it — it’s signal-set-and-weights). Vincent wasn’t aware → the gap is founder awareness/sync, not missing structure. Agreed: ideas enter tables tagged as ideas, promoted to fact only with explicit evidence.
[2026-06-26] (internal prep — Vincent + Elliot, recorded by Vincent) Surfaced a brain-trust gap: querying for the four R’s / measured factors returned inconsistent answers across phrasings, and Elliot couldn’t tell which page was authoritative. No “ADO methodology” page actually exists (canonical four-R home = discoverability-framework); the inconsistency was retrieval/context, not a wrong page. Agreed: critical infra (four-R / score-factor definitions) shouldn’t be silently set by transcript ingestion — the platform + canonical product pages are the source of truth, and ingest must separate truth vs idea vs hypothesis. The aspirational workflow both want: lock a decision in a call → agent extracts → human validates → committed to the wiki. Action: Vincent to find-or-recreate the ADO methodology overview doc.
[2026-06-12] (internal discussion — company-brain working session, Vincent + Elliot) Brain went live: ghostteam-brain repo cloned to each founder’s home top level; agent.md rebase-on-start/push-on-write; harness-agnostic; ingest-call skill proposes wiki updates for human review/commit (no blind writes; ~2–3 calls/day, one review each). Agreed product wiki is the hardest/highest-priority page needing constant human curation (customers ~90% agent-right). Brain-folder integrity safeguard = agent.md content rules + scratch/ redirect + discipline. Skills centralized in the brain (one “create deck” skill for both). No firm agent strategy for 1–2 weeks (Vincent→Claude planning, Elliot→Codex); 24/7 Hermes auto-ingestion is the next phase once stable; Slack-over-Telegram deferred. Open: how to inject brain context into working sessions — Elliot to propose an approach.
[2026-06-05] (internal discussion) Elliot presented the G-Brain-style setup (sources → PR review gate → shared memory → pulled into any coding agent via MCP/CLI); planned to build a first version of the shared memory that day, then deprioritized it in favor of sprint/quarterly planning. Vincent raised Google Drive permissioning as an alternative and advised starting small and avoiding hype-cycle tools. Agreed the durable requirement: memory system with permissions shared across any client. Noted Paperclip resembles the original “ghost team” vision (a team of specialized agents).