# Sextant

> A local-first agent control plane: memory that knows when it is stale and re-verifies itself, plus an orchestrator that acts on its own schedule.

Last verified: 2026-11-09. Markdown version of https://protechtus.com/sextant.html. Product site: https://usesextant.com. Repository opens at release (github.com/protechtus).

## Numbers (verified 2026-11-09)

| | |
|---|---|
| Core tests | 4,000+ (4,241 counted) |
| MCP tools | 30 |
| CI | Linux, macOS, Windows on every push |
| Python | 3.11+ |
| License | Apache-2.0 |
| Status | v0.1.0, first public release, in daily use by its author |
| Models | local by default (Ollama); OpenAI-compatible servers and Anthropic optional |
| Runtime | local, free on one workstation; no cloud dependency required |

## The problem

Every LLM session starts cold. Memory systems that bolt on long-term storage mostly treat confidence as static: stale facts get down-ranked at best and are rarely re-checked, so an agent will confidently repeat something that stopped being true months ago.

## The idea that matters: decay as behavior

Sextant treats a fact's confidence as something that erodes on a half-life curve and is re-checked on a schedule:

    effective_confidence = c0 * 2^(-age / H)

When confidence falls past a threshold the fact is re-verified, retained, surfaced for review, or quarantined and escalated, rather than merely down-ranked. Re-verification sits behind an explicit switch for safety. States: retain, surface, re-verify, quarantine.

## Components (named as in the repository)

- **mcp**: the memory server. Search, write-with-provenance, confirm, temporal queries, project state; a SQLite index with full-text search and confidence bands. 30 tools.
- **brain**: the orchestrator. A service backed by a local model (Qwen3 via Ollama, no cloud API required) holding a SQLite task queue with leases and retries, a worker that drains it, a cron scheduler for nightly hygiene, executor kinds, a test gate and a weekly narrative, plus an escalation router that routes work to stronger models only where they earn it. Runs as a background service; free on local models.
- **web**: the control plane and web UI (HTMX). Status, tickets, tasks, Brain chat, agents and session history.
- **temporal_graph**: the temporal knowledge graph. Tracks facts across time and scores them by text relevance, recency and confidence; feeds context into the Brain before each task.
- **telegram**: multi-channel control. Full mobile read and write access to memory, tasks, tickets and Brain state. A token-budget preflight checks rate-limit headroom before long runs.

Security posture: services bind to loopback and are reached through an authenticated, access-controlled tunnel; secrets live in gitignored environment files.

## Extracted libraries

epistemic-decay (the decay model), inspiration-engine (semantic-entropy self-consistency with a verify-before-promote gate), rpo-guard (backup miss detection). Private, may open; each tested and packaged at v0.1.0.

## Shipped

The MCP memory server, the Brain core with scheduler and worker, the web UI, the Telegram bot, decay as behavior, the temporal graph, token-budget preflight, and the three extracted libraries at v0.1.0.