Your AI forgets everything. Sextant remembers — and knows when it’s no longer sure.

A self-hosted memory and orchestration layer for your AI sessions. Plain markdown you own. Memory that ages, gets re-checked, and says so when it is no longer sure.

View on GitHub →Quickstart ↓

Apache-2.0 · 4,000+ tests, CI on 3 OSes · 30 MCP tools · in daily use by its maker

claude code — session startsextant-mcp · 30 tools

~/work $ claude
✻ sextant · session primed — 47 facts loaded, 2 flagged stale

> what did we decide about the staging database?

● From memory (verified 2026-08-02):
  Staging Postgres lives at 10.0.0.5.
  Migrations run through CI only — never by hand.

⚠ One related fact is stale (91 days unverified):
  "staging still uses self-signed certs"
  Queued for re-verification. Not asserting it.

Illustrative output, not a recording.

The problem

Every session starts from zero. And the memory tools that fix that have two problems of their own: your knowledge lives in someone else’s cloud, and the memory lies — confidently repeating things that stopped being true months ago, because nothing ever checks.

The answer

Instead of confidently wrong, you get honestly unsure.

Illustrative. A fact ages; Sextant re-verifies it. When reality changes and re-verification fails, the fact is flagged — not repeated.

01
Markdown is the source of truth
Your corpus is plain files you can read, grep, edit, and back up. Everything else is derived and rebuildable.
02
Memory that ages
Verification timestamps, decay jobs, reverify loops. Search that answers ‘starved’ instead of guessing.
03
Search that understands meaning
Hybrid keyword + semantic search, with local embeddings through your Ollama by default.
04
Every device, one corpus
Hub/spoke sync with conflict quarantine: conflicts are parked for review instead of overwritten. Secrets and key material never sync.
05
The boring parts included
4,000+ tests, 3-OS CI, auth on every surface, fail-closed defaults, report-first uninstall.
06
Bring your history
Import Obsidian vaults, ChatGPT and Claude exports in one command.

Quickstart

Clone to primed session in ten minutes.

Windows, Linux, macOS. Python 3.11+. Everything binds loopback and fails closed by default.

Full guides:
getting started
self-hosting on a VPS or Tailscale
Docker

$ git clone https://github.com/protechtus/sextant.git && cd sextant
$ python -m venv .venv && . .venv/bin/activate   # Windows: .venv\Scripts\activate
$ python -m pip install -e ".[dev]"
$ sextant-init && sextant-doctor

$ git clone https://github.com/protechtus/sextant.git && cd sextant
$ python3 -c "import secrets; print(f'ATLAS_MCP_TOKEN={secrets.token_urlsafe(32)}'); print(f'ATLAS_WEB_API_KEY={secrets.token_urlsafe(32)}')" > .env   # keys the stack refuses to start without
$ docker compose up -d mcp web
$ docker compose run --rm mcp sextant-init

An honest comparison

Mem0, Zep, and Letta offer managed memory services for product agents, and Mem0 and Letta are open source too. Sextant is not that: it’s the memory layer for your own sessions, under your own roof, with no per-seat pricing and no cloud dependency. If you want a managed API, use them. If you want a second brain you actually own, this is it.

Instead of confidently wrong, you get honestly unsure.

Markdown is the source of truth

Memory that ages

Search that understands meaning

Every device, one corpus

The boring parts included

Bring your history

Clone to primed session in ten minutes.

Follow the voyage