How Smartopol Works | Architecture

§ 01 · short version

If you only read five minutes, read this.

01
One program, always on.
Smartopol runs as a background service on your machine. Terminal, Discord, Telegram, web, and external tools talk to the same brain. One memory, many doors in.
02
Memory Pyramid is the memory system.
Hierarchical memory that organizes itself as you use it. Conversations, channels, users. Local on disk. Fast hybrid recall (keyword + meaning).
03
Context is built in layers.
What never changes (tools, abilities), what changes slowly (who you are), and what changes every turn (this moment's recall) stay separate. That cuts repeated token spend when the provider supports prefix cache.
04
Optional nightly reflection.
When enabled, a bounded job consolidates the day's memory, promotes what mattered, prunes noise, and updates a small profile file. Auditable. Opt-out.
05
Nothing here is a black box by design.
Tool calls go to a local audit log. Reflections write dated receipts. Plugins run sandboxed with declared permissions. You own data, log, and keys.

§ 02 · Memory Pyramid

How memory is structured.

Standalone product path and company details: smartopol.com/memory. Inside the agent, the same architecture is native.

Level N  [Meta-summaries] ← patterns across topics
 │
Level 2  [Domain summaries] ← work, home, projects
 │
Level 1  [Topic summaries] ← one project, one preference cluster
 /│\
Level 0  [Chunks] ← raw conversation fragments

Three operations keep it alive.

ingest

Messages break into searchable fragments with time and channel context.

consolidate

Related fragments group into topics, then higher summaries. Important things climb.

prune

Revisited items stay sharp. Untouched noise compresses into summaries on purpose.

Recall is hybrid.

Exact matches for names, dates, the word you used.
Meaning matches for “that thing last week” even if the phrasing changes.
Both scored together so neither side alone owns the ranking.
All levels at once so a raw line and its summary can arrive together.
Hot path cache for repeated queries.

why this matters

Bigger context windows help, but dumping everything into a prompt is slow, expensive, and often worse than targeted retrieval.

The right few kilobytes beat a million-token dump. Memory Pyramid decides what “the right few kilobytes” are.

§ 03 · context in layers

What changes fast, what changes slowly, what barely changes.

static

Rarely changes.

Tools, stable identity, capabilities. Reused across conversations.

slow

Who you are.

Profile and habits. The agent reads them; it cannot silently rewrite core permissions.

volatile

Rebuilt every turn.

This moment's recall from the pyramid.

Separating layers lets providers reuse stable prefixes when they support caching. You spend less on repeated tokens for the same system context.

§ 04 · optional reflection

Nightly consolidation (off by default).

When you turn it on, the agent reflects once a day at a time you choose (default 04:00 local).

What it reads

Recent conversation across channels. Pyramid promotions and prunes. Current profile snapshot.

What it writes

Bounded profile sections (traits, patterns, shifts) plus a dated receipt you can open and revert.

What it cannot touch

Core identity, tool permissions, program binary, vault keys, other users' memory. Read-only by design where it matters.

Human judgment stays first

Reflection is a tool, not a takeover. You can disable it, read every receipt, and keep control of the profile.

§ 05 · sessions · models · tools

How the agent actually runs work.

Crash-isolated sessions

Each session is its own actor with its own queue. One bad tool call should not freeze the whole service.

Model order you choose

Primary, fallback, local last-resort. Grok, Claude, GPT, Qwen, Kimi, Ollama, OpenAI-compatible endpoints. Failover keeps history. Per-session /model when you need a one-off.

Real tools

Persistent PTY shell, file ops, web search/fetch, scheduling, sub-agents with scoped tools, sandboxed plugins.

Coding agents as hands

Smartopol is the long-lived brain. Grok Build, Claude Code, Qwen Code, Kimi Code, and similar CLIs can act as tools or peers.Grok = model. Grok Build = coding agent product.

§ 06 · security sketch

Safer defaults, not a trust-me speech.

· Encrypted vault for API keys (master password you control)
· Threat scan on memory writes; recalled memory fenced as data, not instructions
· OS-level sandbox for plugins; signed third-party packages
· Command safety checks before shell execution
· Tamper-evident audit log of tool use
· Telemetry off by default

Full inventory: /features. License and source review: /why-closed-source.

One program. No magic.