MOBIUS-RQA — memory-echo governance for assistants with memory

Keep an assistant's own past guesses from coming back as if they were facts.

Assistants with long-term memory share a failure mode this project calls memory echo: an interpretation the model produced on an earlier turn is stored, retrieved later, and then treated as independent, externally-grounded evidence. Left unchecked it yields a self-reinforcing echo chamber and fabricated self-citations — "as I established earlier, X" — where X was never actually established.

This repository is a working, model-independent, near-deterministic mitigation, with an empirical pilot behind it:

• Provenance-tagged memory. Every stored item records its source class (user / assistant-generated / external / system), and the tag is carried into the assembled context, so an assistant-generated item is visibly not independent evidence. (rqa/graph.py, rqa/prompts.py)
• Output-time self-citation verification. Before a response is emitted, every memory reference it makes is checked against the set of nodes actually injected this turn; references to anything not injected are fabricated self-citations and are stripped. The model is not trusted on provenance — the check is deterministic code, not a model self-report. (sanitize_memory_refs in rqa/governor.py)
• Measured. In a pilot (docs/EMPIRICAL_COMPANION_FINDINGS.md), ungoverned raw memory collapses answer-space plurality to ~1/7 of baseline while governed memory recovers it; the output-time verifier removes fabricated self-citations deterministically, and the effect reproduces across six local models in three families (Gemma-4, Llama-3.1, Qwen-3.5/3.6).

If you are building a memory or personalization layer for a chat assistant or agent, those two primitives — provenance tagging and the output-time citation check — are the reusable, liftable part.

Scope: this is a research pilot (small n), not a hardened library. The measurements establish the mechanism; production use wants larger n and your own integration. See docs/EMPIRICAL_COMPANION_FINDINGS.md for the limitations.

The system it lives in: bounded reflective questioning

MOBIUS-RQA (Reflective Questioning Adapter / BRSA) is also a complete reflective questioning assistant — a sibling of the MOBIUS MMV system. It does not primarily answer; it structures tensions, excavates assumptions, and produces deeper questions under explicit governance. The memory governance above is what keeps that reflection honest: its questions stay grounded in what was actually said, not in the assistant's own earlier guesses.

Spec: docs/SPEC_v0_4.md (full integrated spec). Earlier records remain in docs/SPEC_v0_2.md and docs/SPEC_v0_3_DELTA.md. Current implementation: v0.4 RGC conversation/instrument surface + trained adapter path. The code default remains gemma4:12b; the adapter is available through the launcher/Ollama binding pending owner approval.

Architecture (implemented subset)

input
 → Question Graph pre-noticing retrieval   (graph.py, Essentials filter in governor.py)
 → frozen base LLM via Ollama (gemma4:12b) (llm.py, prompts.py)
 → K diverse question candidates           (machine-checked, schema.py)
 → Stage 1: local self-ranking
 → Stage 2: Pinned External Evaluator      (MMV-L gpt-oss-120B via Groq, evaluator.py)
            with Local Degradation         (outage → Stage 1 result, never halts)
 → Governor boundary check (code, not model self-report)
 → output + Graph write-back + selection telemetry

Boundary discipline, the Essentials-like injection filter, append-only graph with audit log, and bounded reflection limits follow the spec exactly.

Setup

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt   # requests + pyyaml (pytest to run tests)

Generation runs on local models via Ollama; the external evaluator needs GROQ_API_KEY (see below). In the MOBIUS monorepo the shared ../venv313 works as-is (bin/rqa wraps it).

Usage

PY=.venv/bin/python               # or ../venv313/bin/python in the MOBIUS monorepo

$PY -m rqa chat                   # conversational REPL: streaming answer +
                                  #   async reflective sidecar (SPEC v0.3)
$PY -m rqa chat "あなたは何者ですか" # one-shot conversation; light inputs skip sidecar
$PY -m rqa check                  # environment health
$PY -m rqa ask "自己理解の層だけを更新すれば、自己更新AIは安定するのではないか?"
                                  # RGC-routed: direct/guided/instrument by weight
$PY -m rqa ask "あなたは何者ですか" # light opener: fast conversation reply
$PY -m rqa ask --instrument "あなたは何者ですか"
                                  # force dashboard for a light opener
$PY -m rqa ask --brief "..."      # question-only view
$PY -m rqa review path/to/doc.md  # document review mode (dogfood target)
$PY -m rqa graph stats
$PY -m rqa graph search "自己理解層"
$PY -m rqa ask --no-evaluator "..."   # force Stage 1 only

Evaluator credentials: GROQ_API_KEY from the environment or ./.env. The evaluator binding itself is pinned in config/evaluator_binding.yaml — changing it is an evaluator swap and requires human approval (spec §6.5).

Tests

.venv/bin/python -m pytest -q tests/

Unit tests are network-free (fake adapter/evaluator). Latest local run: 66 passed. The live path is exercised via rqa check + a real rqa ask.

Training (Stage C / SPEC §13)

Dedicated venv (.venv-train — keeps the shared MOBIUS venv frozen):

PYT=.venv-train/bin/python
$PYT scripts/prepare_training_data.py        # {{SYSTEM_RQA}} -> real prompt, train/val split
$PYT scripts/train_sft.py --smoke            # 4-step VRAM/stack verification
$PYT scripts/train_sft.py                    # full QLoRA SFT -> models/rqa_adapter_v0_1/
$PYT scripts/eval_adapter.py                 # base vs adapter on human-reviewed holdout

Base: google/gemma-4-12B-it (unified multimodal; LoRA targets the language tower only). 4-bit NF4 + rank 16 + gradient checkpointing fits a 16GB RTX 5070 Ti. Corpus: data/sft_phase1_train.jsonl (426; see docs/PHASE1_CORPUS_REPORT.md). sft_phase1_holdout.jsonl and gold_seed.jsonl are evaluation-only — never train on them.

Not implemented yet (by design — spec staging)

• 読み出し2 cross-contradiction check / 読み出し3 graph curation (Stage B/C)
• web_search / local_RAG / document_fetch tools (Stage B)
• ME5 embedding retrieval (Stage B; current backend is char-trigram)

License

• Code & figures: AGPL-3.0-or-later (see LICENSE).
• Documents under docs/ that carry their own license header (the companion paper drafts) are CC BY-NC-SA 4.0, as stated in those files.

Related — the Möbius program

Part of the MOBIUS program — local-first, AGPL:

• mmv — answer-entitlement runtime: decides whether answering is warranted
• rqa — reflective questioning adapter: deepens the question when it is not
• rcgov — reflective context governor: governs what a model may read
• infinity — composite capstone (MMV × RQA) with an OpenAI-compatible API
• tokyo-insight — on-demand civic-RAG engine for 東京都議会 deliberation records (engine + facts only)