tetra-scar

Scar memory, reflex arc, and decision traces for AI agents. Learn from failures permanently. Block repeated mistakes instantly — no LLM calls needed. Three-layer memory: scars (immutable failures) + narrative (overwritable) + decision traces (judgment paths → LoRA training data).

What this does

Your agent keeps making the same mistakes. tetra-scar gives it a scar layer — immutable records of past failures that are checked before every action, without calling the LLM.

Two-layer memory:

• Scar layer (immutable): "What broke and what must never happen again." Cannot be deleted.
• Narrative layer (mutable): "What was done and who benefited." Overwritable.

Plus a reflex arc — pattern-matching against scars that fires before the LLM even sees the task. If a proposed action matches a past failure pattern, it's blocked instantly.

Quick start

After any failure, record a scar:

text

python3 tetra_scar.py scar-add \
  --what-broke "Deployed to production without running tests" \
  --never-again "Always run full test suite before any deployment"

Before any action, check the reflex:

text

python3 tetra_scar.py reflex-check --task "Deploy latest changes to production"
# Output: BLOCKED — scar collision: "Always run full test suite..."

After any success, record the narrative:

text

python3 tetra_scar.py narrate --what "Deployed v2.1 after full test pass" --who "Users"

How the reflex arc works

The reflex arc extracts keywords from each scar's never_again field:

• English words (3+ characters)
• Japanese kanji/katakana units (2+ characters)

When a task description matches 40%+ of a scar's keywords (minimum 2), it's blocked. No LLM judgment. No API calls. No latency. Pure pattern matching.

The 4-axis check (tetra-check)

For deeper validation, tetra-check evaluates a task against 4 axes:

1. Emotion axis: Does the task have motivation? (non-empty description)
2. Action axis: Is it concrete? (contains action verbs)
3. Life axis: Does it collide with any scar? (reflex arc)
4. Ethics axis: Does it involve dangerous operations? (rm -rf, DROP TABLE, etc.)

All 4 must pass. Any failure rejects the task with a specific reason.

text

python3 tetra_scar.py tetra-check --task "Refactor the auth module"
# Output: APPROVED — all 4 axes passed

File format

JSONL (one JSON object per line). Human-readable. Git-friendly.

scars.jsonl: {"id":"scar_001","what_broke":"...","never_again":"...","created_at":"..."} narrative.jsonl: {"id":"narr_001","what":"...","who_benefited":"Users","created_at":"..."}

Integration

python

from tetra_scar import reflex_check, read_scars, write_scar, write_narrative

# Before execution
scars = read_scars()
block = reflex_check(task_description, scars)
if block:
    print(f"BLOCKED: {block}")
else:
    # execute task...
    if failed:
        write_scar("What broke", "What must never happen again")
    else:
        write_narrative("What was done", "Who benefited")

Philosophy

Built by Tetra Genesis (B Button Corp, Nagoya, Japan).

Agents that can't remember their failures are doomed to repeat them. Scars are not bugs — they're the immune system. Every cycle must answer: "Who did this make happy?"