llm-roundtrip.md

Mneme LLM Roundtrip

Mneme now has a wrapper for the first real AI-assisted compile loop.

Script

• scripts/mneme_llm_roundtrip.py

What it does

Runs this flow:

1. ingest raw evidence
2. prepare LLM bundles
3. optionally validate a candidate JSON file
4. optionally materialize validated entries into compiled outputs

Usage

Prepare bundles for one category (currently: projects, systems, decisions, incidents, people, timeline):

./scripts/mneme_llm_roundtrip.py \
  --root /path/to/workspace \
  --category projects \
  --json

Run the full loop with a candidate file:

./scripts/mneme_llm_roundtrip.py \
  --root /path/to/workspace \
  --category projects \
  --candidate ./candidate-projects.json \
  --json

Why this matters

This is the seam where Mneme stops being only deterministic plumbing and starts supporting real judgment:

• deterministic ingest/prep/validation/materialization
• LLM grouping/summarization in the middle

That is the intended architecture.