● Memory + agents / 2024 → now

Memory System.

LLM sessions forget. Every new conversation starts cold, every tool re-asks the same five questions, and last week's context gets compacted into nothing. So I built the memory layer myself: markdown files I can read, a local index every agent calls into, and a continuity layer that lets a fresh session pick up where the last one stopped from real session notes.

Auto-memory

Retell + Twilio

Mortals

MCP ecosystem

Resume

Founder Pres.

Obsidian + QMD

Primary email

Review pass

Claude / Codex

Convex

75K+ sold

SwiftUI CloudKit

Mission Control

Next.js + Vercel

1,350+ locations

Shopify

Session notes

NightShift

Instagram

Founder

GitHub

MCP forks

<60s response

Operator chapters

QMD retrieval

Founder & CEO

121+ skills

Six figures raised

Founder office

Full timeline

5: agent surfaces wired in
4: domain capsules routed
1: index every agent queries
2024: running since

§ WiringSave in, recall out, one store

obsidian-session-notesstructured notes
claude-memsession history
claude-diarycompaction summaries
obsidian-kb-syncpromotes to canon
transcriptionvoice → markdown

Obsidian vaultplain files, on disk

QMDlocal index · one read surface

Claude Code
Codex
Cursor
Gemini
Antigravity

The agent layer that reads it Full build stack

Chat memory is not memory. Files are memory.

§ LoopA fresh session picks up where the last stopped

Capturesession notesthe session writes what it decided
Persist~/rules/canon + memory, on disk
RecallQMDa fresh session reads + queries
Continuewarm start — any harness

§01Outcome

The AI knows my stack on session one.

A fresh session already knows my companies, my people by role, my product line, my brand voice, my coding preferences, the verified facts that override stale assumptions, and what I shipped last. For anything older it queries the vault instead of asking me to re-explain. This portfolio came out of the same system: written from real session notes, not typed from scratch.

§02Where to look next

Component pages.

Obsidian + QMD: the storage graph and local index.
Transcription pipeline: voice into the graph.
Agent infrastructure: the layer that reads this before it acts.
Custom skills: one skill body across every surface.
Custom MCP: the safety-gated write surface.
CodexBar: the Codex fork that reads the same vault.

§03Build stack

What it is built on.

Anthropic Claude / Primary modelclaude-mem / Session captureclaude-diary / Session capture~/rules/ / File-based canonMEMORY.md auto-memory / Per-project recallVERIFIED_FACTS.md / Override layerObsidian session notes / Durable evidenceObsidian KB sync / Canon promotionQMD / Local search indexHindsight / Agent recallResume from session / Continuity skill/reflect / Continuity command/context-handoff / Continuity commandCustom MCP servers / Safety-gated writesClaude Code / Agent surfaceCodex / Agent surfaceCursor / Agent surfaceGemini / Agent surfaceAntigravity / Agent surface

§04Artifacts

The shippable evidence.

~/rules/ : file-based canon

`people.md`, `companies.md`, `projects.md`, `preferences.md`. Plain markdown, symlinked into every harness and loaded at session start, so I never re-explain the operating frame.

private · global canon

Per-project auto-memory

`MEMORY.md` plus `VERIFIED_FACTS.md` and `feedback-*.md` corrections under `~/.claude/projects/<project>/memory/`. Scoped to each codebase, persisted across sessions.

private · per-codebase

claude-mem and claude-diary

Automatic session-history capture across Claude Code runs, plus a structured summary on every /compact. Context fills up without losing the record.

private · session capture

Obsidian session notes

Structured notes written into the right domain capsule (Mortals, LOA, Projects, Personal) at the end of meaningful agent work. The evidence trail every other layer reads against.

private · durable memory

Obsidian KB sync

Promotes durable facts from session notes into canonical domain guides: current manufacturer, current pipeline state, current preferences. Approval-gated so canon doesn't drift on its own.

private · canon promotion

QMD local index

One SQLite-backed full-text and vector index across the whole vault and adjacent local files. CLI plus a localhost MCP server, so every agent tool retrieves through the same index.

private · retrieval layer

Resume from session

Skill that pulls the latest state, next action, and open risks out of recent session notes when a fresh thread continues prior work. The new thread starts where the old one stopped.

private · continuity skill

/reflect and /context-handoff

Pattern analysis across recent sessions, plus an explicit handoff for passing one session's state to the next.

private · continuity commands

§05Adjacent in the grid

Memory System.

The AI knows my stack on session one.

Component pages.

What it is built on.

The shippable evidence.

Keep moving through the cube.