EST. 2026 · OPEN SOURCE · SELF-HOSTED
Modern AI has memory — bound to one product, opaque, single-user, deletable when the vendor changes their mind. MyPA is the archive underneath: structured, queryable, exportable, multi-tenant, and yours. Today Claude reads from it. Tomorrow whatever comes next does too.
Yes — Claude has memory now. ChatGPT too. They're useful for "remember I prefer Python." But that memory lives on their servers, is bound to your subscription, can be silently rewritten by the vendor, doesn't move with you when the next AI gets better, and can't reach outside the chat to remind you about anything. This is convenience without sovereignty. MyPA is the archive at the back of the study — structured, queryable, yours for decades.
The honest question
Native AI memory and a personal archive answer different questions. Both useful. Both can coexist. But the shape of the problem is different.
Great for: "Use Python 3.12 by default." "I prefer brief answers." "I live in London."
Great for: durable life data — decisions, contracts, places, preferences, todos, anything you'll still care about in 5 years.
The model: native memory holds your conversational style; MyPA holds the durable facts of your life. They complement each other. Use both.
A live demonstration
Three real flows — saving a decision, getting a morning push, recalling a past choice — animated in your browser. Not a screen-recording, not a slide deck. Press play, or it'll loop on its own.
What goes in the archive
Six kinds of durable life data that benefit from being structured, queryable, and held across decades — not lost in chat transcripts.
"Bought 100 ABC at £2.10 because Q4 margins expanded and the sector hasn't rotated yet." Append-only by design — five years later when you ask "why did I buy that?", your own past reasoning is verbatim. The vendor can't quietly rewrite it.
Renewal dates, account numbers (never passwords — use a password manager), support phone trees, policy IDs, IBANs. The stuff you Google in a panic. Your AI should already have it. Filter kind:contract end_before:2027 — done.
"Pizza Express on King's Road, rated 5★, visited 20 May." Months later: "What restaurants have I rated 5★ in London?" — Claude opens the drawer and reads. Filter by city, by tag, by date.
"Pizza is so good, help me remember." "Cilantro tastes like soap." "Kepei loves dinosaurs." Casual remarks become structured records. Next time you ask for restaurant ideas or birthday-gift suggestions, your AI already knows the whole family.
"Remind me at 14:00 to pick up Kepei." The push lands on your phone — authenticated, on your own infrastructure, no Slack-style notification overload. Daily morning digest of what's due. Per-user-configurable hour and channel.
One MyPA install serves your family or small team. Each member has their own credentials, their own archive, their own notifications. No cross-tenant leakage — verified by tests. The architecture you wish Notion had at five seats.
A fair comparison
Claude memory and ChatGPT memory are real and useful. They solve a different problem. Here's the line-by-line:
| Claude.ai memory | ChatGPT memory | MyPA + any MCP AI | |
|---|---|---|---|
| Carries context across chats | yes — within one Claude account | yes — within one ChatGPT account | yes — across any MCP-compatible AI |
| Structured queries (kind, date, tag) | no — freeform text only | no — freeform text only | yes — kinds, fields, FTS5 full-text, filters |
| Append-only decision history | no — silently revised by the model | no — silently revised | yes — your past reasoning is verbatim, forever |
| Pushes reminders to your phone | no — chat-bound | no — chat-bound | yes — self-hosted ntfy, authenticated, free |
| Daily morning digest of what's due | no | no | yes — per-user configurable hour |
| Where your data physically lives | Anthropic's cloud | OpenAI's cloud | your VPS — SQLCipher-encrypted, your key |
| Vendor can read your data | yes — for safety + model training opt-out flow | yes — same | no — they never see it |
| Survives if vendor changes ToS / pricing | no — bound to subscription | no — same | yes — you own the install + data |
| Moves with you to the next AI | no — Claude-only | no — ChatGPT-only | yes — MCP is an open protocol |
| Family of 5 (each isolated) | 5 × Pro at ~£15/mo = £75/mo | 5 × Plus at ~£16/mo = £80/mo | 1 VPS at ~£10/mo = £10/mo |
| Exportable to Obsidian/Notion/anything | limited — copy/paste only | limited | yes — Markdown + YAML frontmatter, one command |
| Setup effort | zero — just opens | zero — just opens | ~45 minutes, one time |
Use both. Native AI memory for conversational style and small preferences. MyPA for the durable, structured, queryable facts of your life. They aren't rivals — they're layers.
The compound advantage
Storage on its own is a Notes app. The interesting thing happens when an AI can read it, write to it, search it, push from it, and reason across years of your life — all through one open protocol.
"Look at my last 10 investment decisions — which ones with outcomes recorded actually played out? What's the pattern?" Claude reads every decision with an outcome, compares your thesis to what happened, finds the common thread. Native AI memory can't compare freeform entries across years; structured records can.
→ Decision review, pattern detection, gift-buying retrospectives, "have I gotten better at this?"
The scheduler runs while you sleep. Your AI can be told to think, not just remind: "any contracts renewing in 30 days I should review?", "places I rated 5★ but haven't been back to in a year?", "anyone I haven't logged a call with in 6 weeks?" Surface what you'd want to be told.
→ Contract renewals, relationship maintenance, "you're slipping on X" nudges, weekly retrospectives.
MCP is open. One AI summarises your inbox into notes. Another reads your preferences and drafts gift ideas. A third pulls decisions + outcomes for end-of-quarter review. They share the same archive. Switch vendors any time without migrating data.
→ Email-to-archive pipelines, gift planners, calendar agents, journaling assistants — all on the same data.
"Plan my Tokyo trip — I have places saved there, dentist's on Friday, Kepei has piano Thursday." Your AI reads the archive, drafts the itinerary, schedules the dependencies, writes the emails. Memory becomes the substrate for action, not just retrieval.
→ Trip planning, gift research, weekly schedule drafting, contract-renewal email drafts.
Native AI memory is a sticky note on your monitor. MyPA + an AI is a private analyst who's read every page of your life so far — and can write the next one alongside you.
How it plugs in
MyPA speaks the Model Context Protocol — Anthropic's open standard for connecting tools to LLMs. Claude.ai supports it natively today. Other vendors are following.
Add `https://mypa.<your-domain>/mcp/sse` as a custom connector. OAuth 2.1 with PKCE, family-friendly auth. Mobile inherits desktop's connectors — no second setup.
Settings → Connectors → + Add custom
Add MCP server via the CLI. Same endpoint, same tools (`pa_add`, `pa_search`, `pa_get`, `pa_list`, `pa_delete`, ...). Works from any laptop you SSH from.
claude mcp add mypa <url>
An add-on, not a requirement. If you want phone push for reminders + the morning digest, MyPA ships an integration with self-hosted ntfy — authenticated, per-user topics, free mobile apps. If you don't want notifications, skip this entirely; the archive + Claude integration works fine without it.
ntfy.<your-domain> · or skip
FastAPI REST for scripts/cron, Angular dashboard for browsing. Email + password login per user. Read your data from any browser, anywhere.
mypa.<your-domain>/api/*
Self-host it
Two install paths — pick the one that matches your hardware. Both run the same MyPA, both take well under an hour.
Synology, QNAP, Unraid, TrueNAS, Raspberry Pi, mini-PC, even a spare laptop. Anywhere Docker runs.
git clone https://github.com/kensterinvest/mypa.git
cd mypa/docker
cp .env.example .env
# edit .env: hostnames + admin email + secrets
docker compose up -dBrings up API + MCP + ntfy + Caddy with TLS as four containers, on an internal bridge network. Persistent named volumes. Updates: git pull && docker compose up -d --build.
For a dedicated Ubuntu 22.04 / 24.04 VPS at Hetzner, IONOS, DigitalOcean — anywhere you have root SSH.
git clone https://github.com/kensterinvest/mypa.git /opt/mypa
cd /opt/mypa
sudo bash setup.shInstalls everything system-wide: Python, SQLCipher, ntfy, Caddy, systemd units, Let's Encrypt certificates. Interactive prompts for your hostnames + admin email.
Hardware & DNS
mypa.yourdomain.comntfy.yourdomain.com (optional, only if you want push notifications)setup.sh)Add two A records: mypa.example.com and
ntfy.example.com → your VPS IP. (ntfy is optional —
skip it if you don't want push notifications.)
git clone https://github.com/kensterinvest/mypa.git /opt/mypa
cd /opt/mypa
sudo bash setup.shThe installer pulls dependencies (Caddy, Python, SQLCipher, ntfy), generates secrets, sets up systemd, configures TLS via Let's Encrypt, provisions ntfy with auth on by default. Prompts for your domain, timezone, your admin email.
At the end you'll see your BEARER_TOKEN_RW,
SQLCIPHER_KEY, OAUTH_JWT_SECRET, and the
ntfy publish + admin passwords. Save them in your password manager.
Lose the SQLCipher key, lose the database. (There is no recovery.)
On a desktop browser:
claude.ai → Settings → Connectors → + Add custom →
URL: https://mypa.example.com/mcp/sse → Connect →
login with your email + admin password. Your phone inherits the
connector automatically.
For family members: scripts/add_user.py alice@example.com
creates their account + ntfy login. Share the credentials securely
once.
Honest about security
No magic. The threats MyPA defends against and the ones it doesn't, in plain English.
user_id at three layers, regression-testedauth-default-access: deny-all, authenticated publish onlydisable_user() surgically revokes refresh tokens + ntfy accountOperators get the full threat-model write-up after install (it ships in the install bundle, not in the public repo — see SECURITY.md for vulnerability disclosure).
Frequently answered
Claude memory and ChatGPT memory are great for conversational continuity — "use Python", "I live in London", "I prefer short answers." They aren't great at: structured queries ("all decisions about investments"), preserving the why of past choices without silent revision, surviving a vendor change, multi-user architecture, reaching outside the chat to push a reminder, or being yours when the company gets acquired. Use both — they're different layers of the same need.
For data changes — no. Every Claude tool call hits the live MyPA database in real time. Save an item now → ask Claude about it in the next message → it's there. mypa-api restarts, refresh-token rotation, and new users added on the server are all transparent — Claude auto-reconnects on the next tool call.
Reconnect IS needed only for:
OAUTH_JWT_SECRET — this
invalidates every JWT, so every connector must re-authorize.Claude.ai's mobile app shows notifications for new messages in conversations you're already in — it doesn't have a "wake me at 14:00 with this reminder" API. MyPA's notification layer is a separate concern: a self-hosted ntfy server we wire into the install, with mobile apps that work on iOS + Android. It's optional — if you don't want phone push, skip the ntfy step in setup; the archive itself, Claude integration, and the web dashboard all work without it. We chose ntfy specifically because it's open-source, self-hostable, and doesn't put a third-party notification provider in the middle of your data.
MyPA speaks MCP — Anthropic's open protocol that other vendors are adopting. Switch from Claude.ai to whatever comes next; your archive doesn't move. The new AI plugs into the same MCP endpoint, reads the same data. No migration, no export-import, no lost history. This is the entire point.
£0 software + the cost of a place to run it. If you already have a NAS or a home server, that's free. Otherwise: a £4-15/mo VPS works. No per-user fees — a 5-person family install costs the same as a 1-person install. No API calls to Anthropic are made by MyPA itself (Claude calls Anthropic; MyPA just holds the data).
Yes. The Docker compose stack runs on any Docker
host — Synology Container Manager, QNAP Container Station, Unraid
Community Apps, TrueNAS Scale. Same single-file .env to
configure. Walkthrough: docker/README.md.
Phone photos (typically 5-10 MB at 4032×3024) are downscaled
to a 2048 px longest edge on upload, re-encoded at 85% JPEG
quality, and the EXIF metadata — including GPS coordinates —
is stripped. A 6 MB photo typically becomes ~700 KB. The image is
still high enough resolution to identify everything in the frame, just
not big enough to waste your VPS disk. If you'd rather store originals
exactly as uploaded, set IMAGE_RESIZE_ENABLED=false. The
operator can also tune IMAGE_MAX_DIMENSION +
IMAGE_JPEG_QUALITY.
Five layers, each independently capable of refusing an oversized or abusive upload:
Content-Length header.Content-Type.Every attempt — success or failure — is written to the audit log. All knobs are env-tunable per install.
Pi 4 (2 GB+) or Pi 5 runs MyPA comfortably. Use the Docker path or
the bare-metal setup.sh on Raspberry Pi OS / Ubuntu Server
arm64. For public access without a static IP: Cloudflare Tunnel or a dynamic-DNS provider does the job.
Today: Claude.ai natively (mobile + desktop), plus any agent that speaks MCP. Other vendors are adopting MCP — once their clients support custom connectors, MyPA works without code changes. The MCP protocol is the abstraction.
Your data lives on your VPS, in a SQLite file you can open with standard tools. There's a Markdown-export tool. If MyPA the project goes dormant tomorrow, your install keeps running. Self-hosted means self-sovereign.
No. The dashboard is for browsing. Everything you actually do happens through Claude (or REST scripts). You can skip the dashboard install entirely if you only ever interact via AI.
MyPA is meant for personal-life data — decisions, preferences, life admin. We don't recommend it for regulated data (PII at scale, PHI, financial-services records) without an additional review of your specific compliance regime.
An honest fit-check
The hard part — security model, multi-tenant correctness — is done. The rest is on your hands and your VPS.