mirror of
https://github.com/ruvnet/RuVector.git
synced 2026-07-22 15:43:48 +00:00
edf96d83ed
* feat(mragent): MRAgent graph memory over RuVector with Darwin optimization
Add ADR-269 and a runnable reference implementation of MRAgent ("Memory is
Reconstructed, Not Retrieved") on RuVector, optimized by Meta-Harness Darwin
Mode under the "freeze the model, evolve the harness" invariant.
- Frozen model: deterministic Cue-Tag-Content memory substrate mirroring
RuVector hybrid (RRF) search + bounded-depth Cypher traversal semantics
(examples/mragent/agent/memory.mjs)
- Evolved harness: 10-gene reconstruction genome (cueK, efSearch, hybridAlpha,
fusion, traversalDepth, tagFanout, pruneThreshold, maxContent, rerank,
promptStrategy) in DARWIN_MUTABLE_BLOCK regions (agent/harness.mjs)
- Darwin evolution loop with mapLimit/paretoFront and ADR-150 graceful fallback
when @metaharness/darwin is absent (optimize.mjs)
- scorePolicy.ts fitness mirroring ADR-266; benchmark + probe + 7 deterministic
acceptance gates
- eval corpus with chained multi-hop "bridge" tasks so traversal depth, fan-out
and pruning are genuinely load-bearing
Runs with zero optional deps: baseline 83.3% -> evolved 100% accuracy, faster
and ~33% smaller context. Darwin discovers traversalDepth=3 (LINKED_TO*1..3).
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_017MDmEV4svuFxuDBGg8zek2
* feat(mragent): self-reconstructing graph memory, beyond SOTA (ADR-270)
Extend the MRAgent harness past the paper into calibrated, adaptive,
self-reorganizing memory, co-evolved by Darwin. Also fixes the corpus being
silently excluded by the root .gitignore data/ rule (the example was missing
its eval set).
Beyond-SOTA mechanisms (each a tunable gene Darwin evolves):
- Adaptive depth (haltConfidence): halt traversal once evidence is decisive
- Abstention + risk-adjusted utility (abstainThreshold): refuse on weak
evidence instead of hallucinating; graded on calibrated utility, not raw acc
- Consolidation/replay (agent/consolidate.mjs): store reorganizes its own
topology, laying Cue->shortcut->Content edges (RuVector self-learning GNN)
Substrate upgrades:
- Concept layer (agent/concepts.mjs): dense (concept) vs sparse (token) signals
genuinely decoupled, so hybridAlpha/fusion become load-bearing
- Hardened 24-task corpus, 6 classes (semantic/lexical/hybrid/bridge/
distractor/unanswerable) synthesized from structured signal specs
- All 12 genes proven load-bearing (some via epistatic interaction)
- Memetic optimizer: GA (mapLimit/paretoFront) + multi-start coordinate-descent
polish that reliably finds the narrow calibration optimum
Measured (deterministic, zero optional deps): baseline acc 81% / risk 0.708 /
halluc 0.13 -> evolved 100% / risk 1.000 / halluc 0.00; consolidation -25%
hops at 100% accuracy. 11 acceptance gates pass. ADR-150 compliant.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_017MDmEV4svuFxuDBGg8zek2
* feat(mragent): generalization protocol (train/test/CV) + overfit fixes
Add a held-out evaluation regime that proves the evolved harness GENERALIZES
rather than memorizing the eval set, and fix the overfitting it surfaced.
Protocol:
- Scale corpus to 60 tasks via a deterministic generator (tools/genCorpus.mjs,
npm run gen-corpus), 10 per class, difficulty-varied (1-hop AND 2-hop bridges,
1-3 ranking-distractors) so train constrains every gene
- Optimizer evolves on a class-stratified TRAIN split, selects via 3-fold
cross-validation with a variance penalty (mean - 0.5*range), and reports a
held-out TEST split it never saw
- Generalization gate = does evolution improve the unseen split
Overfit fixes uncovered by held-out eval:
- Abstention confidence now derives from the answer's RAW relevance, not its
decay^depth path score, so deep-but-relevant bridge answers aren't mistaken
for weak ones (b-test confidence 0.39 -> 0.79); abstention generalizes across
depths. Adaptive-depth halt uses the same raw-relevance signal.
- Larger difficulty-varied corpus + CV variance penalty stop the optimizer
shaving under-constrained genes (maxContent->1) to train-fragile settings
Result (held-out test, reproducible): baseline ~30% acc / risk 0.25 / halluc
0.17 -> evolved ~65% / 0.81 / 0.04 (+35pt acc, +0.56 risk). Honest ceiling
(~80%) documented: synthetic embedding noise + one global hybridAlpha can't
serve both dense- and sparse-keyed queries. 12 acceptance gates pass.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_017MDmEV4svuFxuDBGg8zek2
* feat(mragent): GPU LLM write-layer for the Darwin optimizer (local RTX 5080)
Adds the directed-proposal layer the GA lacks (ADR-260 'real Darwin write-layer
proposes leaps from failure traces'): agent/llmMutator.mjs shows a local,
GPU-served code model (qwen2.5-coder via an OpenAI-compatible endpoint) the
current genome + its failing cases and asks for improved genomes. Every proposal
is clamped to the declared gene bounds (coerceGenome) before entering the
population, so untrusted LLM output can only ever be a safe genome — never an
unsafe gene. Wired into optimize.mjs every 3rd generation; folded into the
archive so GPU candidates compete in polish + acceptance.
Fully opt-in + gracefully degrading (ADR-150): MRAGENT_LLM=off or no reachable
endpoint => identical deterministic GA+coordinate-descent run as before. Auto-
detects http://localhost:11434/v1 (ollama) by default; MRAGENT_LLM_URL/MODEL
override.
Measured (RTX 5080, qwen2.5-coder:7b): 8 genomes proposed across gens, bounds-
safe; the deterministic polish still wins on this small synthetic corpus (the
GA+grid already enumerates the optimum), so the write-layer is a no-regression
enhancement that matters on larger corpora the grid can't cover. 14/14 tests
pass (2 new coerceGenome safety tests).
Co-Authored-By: claude-flow <ruv@ruv.net>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: ruvnet <ruvnet@gmail.com>
|
||
|---|---|---|
| .. | ||
| genCorpus.mjs | ||