mora

README

Mora

Local memory for MCP agents.

Mora backfills your Gmail, Google Calendar, iMessage, Apple Calendar, and local files into a vault of plain Markdown files plus a SQLite index on your machine, and serves it over MCP to Claude Code, Codex, or any other MCP client. Agents answer from your actual history — people, commitments, decisions — with citations. There is no server, account, or telemetry; the only network connections are to the sources you sync, GitHub during mora upgrade, an optional localhost Ollama embedder, and — only if you opt in — a private git remote you control for vault backup (mora sync git).

What it looks like

$ mora think "what did we decide with Sam about pricing?"

Evidence (3):
  [mem_20260610_204655_f3049131] Pricing call with Sam — Sam agreed to $29 one-time
  for the pilot; revisit a subscription tier once we cross 10 seats. He wants the
  invoice before Friday.
  …

Gaps: none detected.

think returns cited evidence plus gaps — stale results, thin coverage, a name the vault has never seen — computed before any model runs. Mora has no LLM and no API key; the calling agent writes the answer, and over MCP it calls search_memory, think, and brief itself.

Install

curl -fsSL https://raw.githubusercontent.com/pyranthus-hq/mora/main/install.sh | sh

The script installs the release binary for your platform, clears macOS Gatekeeper quarantine (binaries are ad-hoc signed, not notarized), and runs mora init (vault at ~/vault/mora). It does not verify checksums — they're on each release if you want to check manually, or unpack a tarball yourself. From source: go install github.com/pyranthus-hq/mora/cmd/mora@latest (Go 1.22+, no CGO; source builds report dev and cannot self-update).

Then connect sources and wire in an agent:

mora connect google                    # OAuth → Gmail + Calendar backfill (read-only scopes)
mora connect imessage                  # macOS; walks you through Full Disk Access
mora schedule install ingest-hourly    # background sync (launchd; prints a cron line on Linux)

claude mcp add mora -s user -- mora mcp serve    # Claude Code
codex  mcp add mora -- mora mcp serve            # Codex

Per-connector setup, options, and upgrades: the guide.

Why a local corpus

Claude and ChatGPT connectors fetch from Google's APIs per query and process results server-side — good for "what's on my calendar tomorrow," not for "what did I commit to, and to whom" across months. Mora keeps a persistent corpus instead, and indexes iMessage, which has no cloud API.

	Mora	Cloud connectors	MCP Gmail servers
Data lives	Your disk (Markdown + SQLite)	Vendor cloud	Nowhere — live fetch
iMessage	Yes (local chat.db, read-only)	No	No
One person across email + texts + calendar	Yes	No	No
Works offline / greppable	Yes	No	No
Cost	$0, no account	Subscription	Free

Cloud tools win on zero setup, a web UI, and write actions; Mora has none of those. Sources and caveats: the guide.

What you get

• Your history as plain files you own. Every email thread, text conversation, and calendar event becomes one Markdown file on your disk — open them in Obsidian, grep them, back them up like any folder. The search database is just an index: delete it and mora index rebuild recreates it from the files.
• Your documents, searchable. Point Mora at a folder and it indexes your notes and text files and pulls the words out of Word docs and PDFs — and a PDF someone texts you over iMessage becomes a searchable memory of its own. Mora only indexes text it can actually read: a scanned, image-only PDF yields nothing rather than garbage. OCR it yourself and drop the text into a files source.
• One view of each person. mora graph "Sam" gathers a person across email, texts, and calendar — shared threads, meetings, the people they show up with — and every claim cites the memory it came from. Built with conservative rules, not an AI model: it never guesses that two people are the same, and no-reply robots are filed as services, not people.
• Search that gets what you meant. Keyword search works out of the box. For "find the thing I'm describing, not the exact words I used," mora config embedder ollama turns on semantic search via Ollama — still entirely on your machine. How retrieval works.
• A morning briefing. mora brief: what's new, what you haven't answered, what's coming up — ranked by who actually matters to you, not whoever emailed last.
• A backup you control (opt-in). mora sync git pushes the vault to a private git remote you choose — GitHub, GitLab, self-hosted, even a bare repo on a USB drive. It only ever pushes, fails loudly rather than guessing, and never includes the index or your login tokens. Restore is git clone + mora index rebuild. Details.
• Plugs into your AI agent. 11 MCP tools let Claude Code, Codex, or any MCP-capable agent search your memory, read the brief, and save new facts back (write_memory). Every search answer says how fresh each source is, so an agent can't silently lean on stale data. Wiring guide.

Privacy model

• Read-only. Google scopes are gmail.readonly and calendar.readonly; the iMessage and Apple Calendar databases are opened read-only. Mora cannot send, modify, or delete anything.
• All data local. Vault, index, and OAuth tokens (~/.config/mora/tokens/, 0600) stay on disk. No analytics endpoint; the local, content-free usage log is disabled by mora usage off or DO_NOT_TRACK=1.
• Zero egress by default. Mora runs no server and never hosts your data. The one opt-in exception is mora sync git, which pushes the vault to a private remote you control — and mora doctor warns whenever the vault is a git repo. For ciphertext at the remote, layer git-remote-gcrypt over it.

Docs

The guide — connectors, MCP wiring, daily use, how retrieval works, the cloud comparison. docs/architecture/ — contributor spec: 13 subsystem docs with diagrams and file:line citations.

---

_{Named for Hermaeus Mora, keeper of knowledge and memory.}