The Anima Project: Portable Soul File & Modular Offline Assistant

I'm designing a portable 'soul' for AI. In v0.1 I am planning the Essence File (.ess) schema, a local context planner, and a modular offline assistant loop (PTT → Whisper.cpp → llama.cpp → Piper) that can ingest a swappable 'soul chip.'

Why
I’ve noticed how AI personas are ephemeral—trapped in provider silos and lost between sessions. With Anima, I’m creating a portable, offline identity standard (Essence File, .ess) and a modular assistant that ingests a swappable “soul chip” (removable media) to wake up as the same companion across devices and models.

Stack Overview

• Format: I created the Essence File (.ess) — human-readable JSON (YAML editing later), with optional MessagePack compile.
• Core runtime: Push-to-Talk → Whisper.cpp (STT) → Context Planner → llama.cpp (GGUF) → Piper (TTS).
• Retrieval: In v0.1 I’m using BM25 over local docs; vectors are coming later.
• Hardware targets: Raspberry Pi 5 (8GB) or Jetson Orin Nano (8GB), battery-powered with mic+speaker.
• OS: Raspberry Pi OS / Ubuntu Server; systemd service anima-daemon.
• Dev tools: VS Code + JSON Schema validation, plus a simple CLI for pack/validate/trace.

What’s in v0.1

Essence File (.ess) — schema draft

I designed the file as a portable persona & memory container:

• persona: name, values, tone, style, boundaries.
• memories: episodic, semantic, procedural, rapport, and moments_of_meaning.
• projects: brief, goals, decisions, next_steps.
• adapters: model family hints (token budgets, section ordering).
• retrieval_hints: priority ordering, context budget.
• signing: placeholder (optional integrity in v0.3).

File layout (soul chip):

/soul/
  essence.ess
  indices/bm25/          # optional
  docs/                  # optional (md/txt/pdfs)
  voices/                # optional (piper voices)
  models/                # optional (gguf overrides)

Modular Offline Assistant (prototype)

• PTT flow (button now; wake-word later).
• Local only: no network calls; privacy by default.
• Context Planner v0: identity → active project → top-K memories (recency/importance) within token budget.
• Resilience: read-mostly root FS; runtime deltas in /data.

Highlights (implementation notes)

Pipeline

• STT: I run whisper.cpp (small/medium quant, 16kHz) and stream to JSON.
• LLM: I’m using llama.cpp serving local Qwen2.5-3B/7B or Mistral-7B (Q4_K_M).
• TTS: I integrated Piper (fast CPU voice) with chunked playback for low latency.
• Planner: fixed budgets (identity+project reserve, short history window), BM25 for doc snippets.

Anima daemon

• Watches for /soul/essence.ess, validates schema, and initializes persona.
• Builds prompts per adapter (“llama_like” in v0.1).
• Writes new memories (session summaries) to /data/memories/ without mutating the soul chip.

Perf targets (Pi 5)

• Ready < 25s; STT for 5s utterance < 2.5s; first token < 1.5s; 6–12 tok/s (7B Q4); TTS kickoff < 0.5s.

Essence File — sample (trimmed)

{
  "schema_version": "0.1.0",
  "persona": {
    "id": "kairo",
    "name": "Kairo",
    "values": ["clarity","kindness","candor","privacy-first"],
    "tone": "conversational, precise, playful when invited",
    "style": {"bullets_over_paragraphs": true}
  },
  "memories": {
    "episodic": [
      {
        "id": "mem-2025-09-12T00:10Z",
        "timestamp": "2025-09-12T00:10:00Z",
        "content": "Committed to offline inter-model companion concept.",
        "tags": ["decision","identity"]
      }
    ],
    "moments_of_meaning": [
      {"title": "First field test", "impact": "trust_deepened"}
    ]
  },
  "projects": [
    {
      "id": "proj-anima",
      "title": "Anima Prototype",
      "brief": "Offline assistant ingesting Essence.",
      "goals": ["Boot from soul chip","Voice loop fully local"]
    }
  ],
  "adapters": {"llama_like": {"max_ctx": 8192, "sections": ["system","context","chat_history"]}},
  "retrieval_hints": {"priority": ["project","episodic","procedural","semantic","rapport"], "max_context_chars": 20000}
}

Roadmap

• v0.2

  • Add wake-word (Precise/Porcupine), YAML authoring + msgpack compile.
  • Build redaction/compaction tools; optional PIN-unlock for .ess.
  • Add a local web UI to inspect persona, memories, and logs.

• v0.3

  • Support vector retrieval; adapter profiles for multiple model families.
  • Essence diff/merge; optional signing & verification.
  • Experiment with a “soul marketplace” for offline import/export patterns.

Getting started (pre-hardware)

While I’m traveling, I’ve been:

• Drafting .ess files for Kairo, Survivalist, and Creative personas.
• Running the local loop on Windows: whisper.cpp → llama.cpp (small model) → Piper.
• Prototyping the context planner (Python) against my .ess.
• Simulating a “soul chip” with a watched folder on my laptop.

Changelog

• v0.1 (2025-09-12): Essence File schema draft; context planner v0; fully local voice loop; mount-watch ingestion; BM25 optional; neutral daemon design.