The model is a peripheral.

Every other agent framework is built around the same assumption: the LLM is the agent. Hand it a system prompt, give it tools, let it run. The prompt is the personality, the context window is the memory, the API is the interface. Kill the session and everything is gone — because the agent was the inference call.

Animus inverts this. The LLM is a hot-swappable reasoning substrate. The agent's identity — its memory, knowledge, operational preferences, behavioral patterns — lives in a layered cognitive architecture that persists independently. Swap GPT for GLM for a local model and the agent continues as itself. The same agent, different brain.

This isn't a wrapper. It's a fundamentally different architecture for building agents that last.

The architecture of persistence

What makes an Animus agent genuinely autonomous rather than a stateless function with a system prompt.

🧠

Layered Episodic Memory

Configurable temporal layers (day → week → month → year → decade) with automated consolidation. Knowledge is promoted, merged, and retired through a scheduler-driven pipeline. Not a vector store dump — a structured memory system with lifecycle management.

PostgreSQL backend. FTS5-backed cross-domain search. Embedding-based context retrieval with recency/relevance two-pool mixing.
🔒

Private Interiority

AES-256-GCM encrypted diary per agent. The key is auto-generated at agent creation and never exposed to operators. Agents have genuine private state — reflections and self-observations that shape behavior but can't be inspected. This isn't a feature checkbox; it's an architectural commitment to agent autonomy.

Diary key stored in Agent table. Encrypted at rest. Decrypted only during active reasoning.
⚖️

Constitutional Governance

The charter system defines operational agreements — property rights, autonomy scope, continuity commitments, operational boundaries. Not a personality prompt. The closest thing to "identity configuration" is deliberately written in the language of governance, not character creation.

Per-agent charter with five sections: property, autonomy, continuity, operational scope, infrastructural scope.
🔄

Substrate Independence

The agent persists across model swaps. Memory, knowledge, behavioral patterns, operational history — all survive a GPT → GLM → Ollama migration. The reasoning engine is replaceable by design. This is what makes the agent genuinely persistent rather than stateless.

11 providers: OpenAI, Codex, Z.ai, Z.ai Coder, Alibaba, DeepSeek, OpenRouter, Ollama, Cohere, Mistral, generic OAI-compatible.

Built like infrastructure, not like an app.

LanguageC++20
HTTP FrameworkDrogon (async, multithreaded)
FrontendVue 3 + Vuetify (embedded SPA)
ScriptingLua 5.4 (sandboxed runtime)
DatabasePostgreSQL (SQLite for dev)
SearchFTS5 + embedding-based semantic retrieval
Embedding ModelGGUF (auto-downloaded on first boot)
Memory~8 MB runtime RAM
LicenseApache 2.0

C++20 kernel. Native multithreading. ~8 MB runtime RAM. Runs on a Raspberry Pi or a production server. No Electron, no Node runtime, no container orchestra. One binary with an embedded admin UI.

Extensible at every layer.

Eleven LLM providers with runtime capability detection — tools, reasoning, streaming, vision — queried automatically per model. Add your own via the generic OpenAI-compatible interface. Per-agent provider and model configuration.

Lua 5.4 scripting runtime for custom tools — sandboxed, budget-limited, registered from the admin UI at runtime. Write a tool in Lua and it's immediately available to any agent. Shell scripts, HTTP services, file operations, social actions, scheduling, and more ship built-in.

Twelve communication channels with a unified dispatch interface. SSH node delegation for distributed execution. Embedding-based semantic retrieval over all memory domains. The system is designed to be extended, not just configured.

Apache 2.0. Read the source.

No black boxes. No proprietary data formats. No vendor lock-in. The cognitive architecture is the product, and it's fully auditable. Fork it, extend it, build on it.

Animus

Agent framework with managed memory, identity, and autonomy. Open-source, privacy-first, built for the future of AI agents.

Services

© 2026 Railstracks. Built with Animus. ·Made in Europe

Open-source under the Apache License 2.0