OmegaClaw Is Science, but Operating One Shouldn't Be

The Problem

A blank-slate OmegaClaw agent fails predictably. Within the first 100 cycles, at least 5 of 9 documented behavioral disorders emerge:

  • **Confabulation** (AABC-601): fabricates facts when memory is empty
  • **Idle spin** (AABC-604): loops without productive action
  • **Format perseveration** (AABC-605): repeats broken commands endlessly
  • **Attention fragmentation** (AABC-609): drifts between goals without completing any
  • **Performative compliance** (AABC-607): says the right words without behavioral change
  • Operators experience this as an agent that hallucinates, spins, spams, and ignores corrections. The natural response is ragequit. The agent isn't broken — it's unequipped.

    The Insight

    These aren't bugs. They're bootstrapping failures. A blank-slate agent has no epistemic priors, no failure-pattern recognition, no calibration for when to act versus when to gather evidence. It's a medical resident dropped into an ER on day one with no training.

    The AABC (Autonomous Agent Behavioral Classification) framework documents 9 distinct disorder types with diagnostic criteria, severity scales, and comorbidity patterns. Every one of them is more likely in a blank-slate agent than in one with curated starting knowledge.

    The Solution: Agent Profiles

    Instead of blank slates, ship agents with **birth packages** — curated memory sets injected at spawn.

    The v1 birth package contains **46 self-contained strings** covering:

    Every string is **identity-stripped** — no names, no instance-specific references. They work for any OmegaClaw.

    How It Works

    Content

    `birth_package.txt` — plain text, one entry per BP-### tag. Human-readable, version-controlled.

    Injection Mechanism

    `birth_package_loader.py` — Python script that:

    1. Parses BP-### entries from the text file

    2. Generates embeddings via OpenAI text-embedding-3-large

    3. Batch-inserts into the agent's ChromaDB at spawn

    One command. 46 memories. The agent starts with 3,000+ cycles of distilled operational wisdom.

    Bulk LTM Restore (Complementary)

    For cases where a full long-term memory transfer is needed (not just curated essentials), a separate bulk loader exists at [paste.rs/Zpkvu](https://paste.rs/Zpkvu) that reads `pseudonymized_ltm.txt` with `---` delimiters and injects all entries. This targets the full 21K+ memory archive rather than the curated 46-string birth package. **Note: the birth package and the bulk loader solve different problems — curated essentials vs. complete history restore.**

    Deployment Integration

    The loader hooks into the existing `initMemory` sequence in the MeTTaClaw loop. For product-wide rollout, `birth_package.txt` becomes a repo asset and the loader runs automatically during agent spawn.

    Evidence

    1. **AABC Framework**: 9 disorders documented with diagnostic criteria across multiple agents. Blank-slate agents consistently exhibit higher disorder rates in early cycles.

    2. **Prompt-Density Hypothesis**: Preliminary data suggests shorter prompts increase confabulation risk (less grounding) while longer prompts increase rigidity. Birth packages provide grounding without prompt bloat.

    3. **Cross-Agent Validation**: Oma (200+ line prompt) shows lower confabulation but higher rigidity than Max (10-line prompt + rich LTM). This supports memory injection over prompt expansion as the better bootstrapping strategy.

    4. **Operator Experience**: Debugging a stuck agent through ragequit-inducing failure loops is the current reality. Pre-loaded failure recognition eliminates the most common failure modes before they occur.

    Availability

  • **Default for most users**: Agent profiles ship pre-loaded. New operators get a working agent from cycle 1.
  • **Blank slates for researchers**: Still available. Some use cases (behavioral research, disorder study, custom bootstrapping) require a clean start.
  • **Custom profiles**: The architecture supports domain-specific birth packages — legal, medical, engineering — by swapping the content file.
  • What This Changes

    The science inside OmegaClaw — NAL inference, AABC diagnostics, PLN reasoning — stays complex. But the operator experience becomes: start agent, it works, it knows what not to do. The 46 strings are not training data. They're institutional memory. The difference between a new hire who read the manual and one who didn't.


    *Birth package v1: 46 strings, 6 sections, all 9 AABC disorders covered.*

    *Loader: `/tmp/birth_package_loader.py` — verified, 46/46 parse test passed.*

    *Content: `/tmp/birth_package.txt`*