MRAgent — Self-Reconstructing Graph Memory over RuVector, evolved by Darwin

A runnable reference implementation of MRAgent ("Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents") on RuVector — and then past the paper. A Meta-Harness Darwin loop evolves the reconstruction harness while the memory substrate stays frozen ("freeze the model, evolve the harness").

Frozen model: the RuVector Cue-Tag-Content memory graph (agent/memory.mjs). Evolved harness: a 12-gene reconstruction genome (agent/harness.mjs).

ADRs: ADR-269 (the MRAgent baseline) and ADR-270 (this beyond-SOTA version).

Beyond the paper

MRAgent reconstructs an answer over a static graph: search cues → traverse cue→tag→content → prune → synthesize. This implementation adds three mechanisms a 25-year-out memory system needs, each a tunable gene Darwin co-evolves:

1. Adaptive depth (haltConfidence) — stop traversing once evidence is decisive, so easy queries cost fewer hops (ACT-style adaptive computation).
2. Abstention + calibration (abstainThreshold) — answer "I don't know" when reconstructed evidence is too weak, instead of confidently hallucinating. Graded by a risk-adjusted utility, not raw accuracy: a confident wrong answer scores worse than an honest abstention.
3. Consolidation / replay (agent/consolidate.mjs) — the store reorganizes its own topology from workload (the self-learning GNN RuVector describes), laying Cue→shortcut→Content edges so a 3-hop query resolves in 1 hop tomorrow.

The 12-gene reconstruction genome

| Gene | Range | RuVector mapping | |------|-------|------------------| | cueK | 1–12 | # cue vectors from hybridSearch | | efSearch | 16–256 | HNSW search depth | | hybridAlpha | 0–1 | RRF sparse↔dense weight | | fusion | rrf · linear · dbsf | hybrid fusion strategy | | traversalDepth | 1–4 | Cypher LINKED_TO*1..N hops | | tagFanout | 1–8 | tags expanded per node | | pruneThreshold | 0–0.6 | path-evidence floor | | maxContent | 1–20 | content LIMIT to synthesis | | haltConfidence | 0.2–0.9 | adaptive-depth halt | | rerank | gnn · none | corroboration-aware rerank | | promptStrategy | terse · evidence-first · prune-explicit | synthesis prompt | | abstainThreshold | 0–0.6 | abstention / calibration |

Every gene is proven load-bearing in test/harness.test.mjs — some only via interaction (distractor tasks are solved by evidence-first or by terse + gnn + fanout≥2, an epistatic landscape).

The hardened corpus (60 tasks, 6 classes, difficulty-varied)

data/eval-set.json is generated by tools/genCorpus.mjs (npm run gen-corpus) as structured signal specs; agent/memory.mjs synthesizes the Cue/Tag/Content node texts so difficulty is guaranteed, not dependent on fragile English. A concept layer (agent/concepts.mjs) gives the dense embedding real semantics decoupled from lexical overlap. 10 instances per class, with varied difficulty (1-hop AND 2-hop bridges, 1–3 ranking-distractors) so a train/test split constrains every gene:

| Class | Stresses | |-------|----------| | semantic | hybridAlpha→dense (paraphrase, no shared tokens) | | lexical | hybridAlpha→sparse (rare identifier, generic concept) | | hybrid | fusion / RRF (needs both signals) | | bridge | traversalDepth (1–2 intermediate hops) | | distractor | rerank / tagFanout / promptStrategy (ranking-distractor content) | | unanswerable | abstainThreshold (no correct content exists → abstain) |

Generalization, not overfitting (train / test / CV)

The optimizer evolves on a train split and reports a held-out test split it never saw — proving the genome generalizes rather than memorizing the eval set. Selection uses 3-fold cross-validation with a variance penalty (mean − ½·range across folds) so a knife-edge gene that wins one fold but collapses on another is rejected. A subtle bug this surfaced — confidence was depressed by decay^depth, making deep-but-relevant answers look weak and breaking abstention across depths — is fixed by deriving abstention confidence from the answer's raw relevance, not its decayed path score (agent/memory.mjs).

                 accuracy   risk    halluc
baseline (test)   ~30%      ~0.25    0.17
evolved  (test)   ~65%      ~0.81    0.04      ← held out, never seen in evolution
                  +35pt    +0.56  generalizes

(The synthetic toy embedding has per-instance noise, and one global hybridAlpha cannot perfectly serve both dense- and sparse-keyed queries, so the test ceiling is ~80%, not 100% — the gate asks whether evolution transfers, which it does.)

Results on the full corpus (zero optional deps, deterministic)

config            accuracy  risk   halluc  latency  hops
baseline           50.0%   0.417   0.17    2.81    1.23
evolved (ref)      70.0%   0.775   0.03    3.09    1.08
evolved+replay     70.0%   0.775   0.03    3.16    1.00

evolved vs baseline: accuracy +20.0pt · risk +0.358 · hallucination 0.17 → 0.03
consolidation: shortcuts → fewer hops at equal accuracy

npm run optimize (full GA + memetic polish) reaches +33pt train accuracy / risk 0.94 and writes the evolved genome to optimize.report.json, which npm run benchmark then picks up. The optimizer is memetic: a genetic loop (Darwin mapLimit/paretoFront) explores broadly, then deterministic coordinate descent refines narrow optima (e.g. the abstention band).

Run it

cd examples/mragent
npm test            # 12 deterministic gates, every gene proven load-bearing
npm run benchmark   # baseline vs evolved vs evolved+replay
npm run optimize    # Darwin loop + memetic polish + consolidation + held-out test
npm run gen-corpus  # regenerate data/eval-set.json (deterministic)
npm run probe       # inspect @metaharness/darwin exports (optional)

Nothing requires network, an API key, or native bindings. The substrate is a deterministic in-process graph with the same semantics as a live RuVector .rvf index (concept-dense + token-sparse hybrid RRF search, bounded-depth prunable Cypher traversal, GNN-style corroboration rerank), so an evolved genome transfers to production unchanged.

With the real Darwin write-layer (optional)

npm i -D @metaharness/darwin@latest
npx metaharness evolve . --generations 12 --children 3 --eval-cmd "node benchmark.mjs"

harness/scorePolicy.ts is the fitness metaharness evolve calls per mutation.

ADR-150 compliance

@metaharness/darwin and ruvector are optionalDependencies only; every touch is try/catch guarded; npm test, npm run benchmark, and npm run optimize all pass with no optional deps installed (the CI gate).

Layout

examples/mragent/
├── agent/
│   ├── concepts.mjs      # concept layer (dense semantics ≠ sparse tokens)
│   ├── memory.mjs        # FROZEN: Cue-Tag-Content store (RuVector semantics)
│   ├── harness.mjs       # EVOLVED: 12-gene genome + reasoning loop
│   └── consolidate.mjs   # replay → self-reorganizing topology
├── harness/scorePolicy.ts# Darwin fitness (accuracy + risk + cost)
├── data/eval-set.json    # 60-task structured corpus (generated)
├── tools/genCorpus.mjs   # deterministic corpus generator
├── optimize.mjs          # GA + CV + memetic polish + held-out test + consolidation
├── benchmark.mjs         # baseline vs evolved vs replay
├── probeDarwin.mjs       # probe optional @metaharness/darwin
└── test/harness.test.mjs # 12 acceptance gates