NeuralMind v3.9.0 — SOTA Synapse Dynamics & Adversarial Retrieval

Release Date: September 2026

TL;DR

Six modern brain-inspired learning techniques in the synapse layer, plus three adversarial fixes for shallow retrieval. The synapse layer now learns like a biological brain — not just “what files go together” but “what you actually need before you ask.”

• Synapse Dynamics (6 techniques): Lateral inhibition, Synaptic Tagging & Capture (STC), non-monotonic plasticity (SAMPL), resource-dependent STDP, Feeling-of-Knowing (FOK) gating, replay-based consolidation
• Retrieval Enhancements (3 fixes): Intent-aware classification, code-signal boosting, synapse-seeded expansion, dependency graph traversal, code snippet extraction
• Token reduction: 32.6x (1,534 tokens vs 50K full codebase)
• Test coverage: 54 new tests, all passing

Synapse Dynamics

The synapse layer now implements six modern computational neuroscience techniques:

1. Lateral Inhibition (SYNAPSE paper, arXiv 2025)

When one concept activates, it suppresses competing activations rather than only boosting neighbors. Prevents “attention dilution” in large codebases.

2. Synaptic Tagging & Capture (PNAS Nexus 2022)

Not every co-activation is meaningful. Two-phase model:

• Tag: Co-activation creates a temporary mark (short-term)
• Capture: If the same pair fires again within a consolidation window, tagged synapses become permanent
• Decay: Untagged marks fade without entering long-term memory

3. Non-Monotonic Plasticity (SAMPL model, bioRxiv)

Memory retrieval both enhances the retrieved item AND weakens related but non-retrieved items. Prevents “everything is vaguely associated with everything.”

4. Resource-Dependent Heterosynaptic STDP (Frontiers 2025)

Each node has a finite local resource pool. Strengthening edge (A,B) consumes resources from A’s pool, naturally weakening competing edges (A,C), (A,D). Creates synaptic competition without global normalization.

5. Feeling-of-Knowing (FOK) Gating (SYNAPSE paper)

Confidence gate on retrieval. If peak activation after spreading doesn’t exceed an adaptive threshold, returns empty rather than weakly-associated noise. Prevents hallucination of irrelevant context.

6. Replay-Based Consolidation (bioRxiv 2025)

Replay queue captures recent co-activation sequences. During idle periods, replays them to strengthen associations without new input. Interleaves recent + old patterns to prevent catastrophic forgetting.

Retrieval Enhancements

Problem

For “how does X implement Y” queries, the system was surfacing docstrings instead of implementation code. Qwen 3.8 Flash adversarial QA identified this as the critical failure mode.

Fix 1: Intent-Aware Classification

Queries matching “how does implement/perform/do " are now classified as `code` intent, not `docs`. This ensures implementation queries surface code files.

Fix 2: Code-Signal Boost

Extracts CamelCase, snake_case, and plain identifiers from queries. Boosts results from files whose source code contains those identifiers. Penalizes docs (0.3x multiplier) for code-intent queries.

Fix 3: Synapse-Seeded Expansion

Checks if query terms match known synapse nodes. If so, spreads activation through the synapse graph to find co-implemented neighbors.

Fix 4: Dependency Graph Traversal

For each identifier, finds its callers/callees/imports in the structural graph. Surfaces code that is structurally related even if not semantically similar.

Fix 5: Code Snippet Extraction

For source file matches, extracts actual source code centered on the best-matching identifier (with line numbers) instead of generic document snippets.

Correction (v3.10.0): fixes 3–5 are off by default.

As first shipped, fixes 3–5 — the passes that pull in nodes vector search did not return — regressed the faithfulness gate and blocked this release. They appended up to eight nodes to a four-hit list and forced their scores to a hardcoded 4.5, so injected nodes outranked every real hit and the genuine ones were pushed past the L3 token budget. On the reference fixture the faithfulness delta went +0.041 → -0.065, under the gate’s +0.000 floor.

They are now behind NEURALMIND_RETRIEVAL_EXPANSION=1, off by default, and budget-neutral when enabled. Making them budget-neutral did not rescue them — displaced, the delta was -0.107, worse than appending — because displacement evicts a real hit per candidate, so candidates worse than what they replace cost facts rather than only tokens. That is a candidate-quality problem, and it is unsolved.

Fixes 1 and 2 (intent classification, code-signal boost) are unaffected and stay on. They re-rank the hits retrieval already returned rather than adding to them, and measure bit-for-bit neutral on the same fixture. The headline “how does X implement Y” behaviour is Fix 1, so it still ships.

See NEURALMIND_RETRIEVAL_EXPANSION in the CLI Reference.

API

from neuralmind import NeuralMind

nm = NeuralMind('/path/to/project')

# Synapse dynamics
nm.dynamics_reinforce(['entity:id:1', 'entity:id:2'])  # STC + resource STDP + replay
related = nm.dynamics_spread([('entity:id:1', 1.0)])    # Lateral inhibition + FOK
stats = nm.dynamics_stats()                             # Introspection

# Enhanced retrieval (automatic)
result = nm.query('How does the synapse layer work?')
# Now surfaces implementation code, not just docstrings

Configuration

Benchmarks

Known Limitations

1. synapses.py surfacing: For “how does X work” queries, synapses.py still ranks below docstrings. The code-signal boost (10x cap) is insufficient to overcome the initial vector score gap. Two-pass retrieval with snippet extraction was the attempted fix and is the pass now disabled by default (see the correction above) — it surfaced the implementation by outranking everything, including the hits carrying the answer. The open problem is a pull-in whose candidates beat the hits they displace; substring matching over identifiers does not.

2. No re-ranking model: Cursor uses a fine-tuned 7B CodeLlama reranker. We use multiplicative score boosting. This is the single biggest quality gap.

3. No dynamic context discovery: Cursor v1.8.3+ writes large tool outputs to files and lets the agent read on demand (46.9% token reduction). We dump all retrieved context into the prompt at once.

4. No independent benchmark: Our 32.6x reduction is self-measured. No SWE-bench or Terminal-Bench score yet.

Migration

No migration needed. Existing synapses.db files are compatible. New tables are created automatically on first use.

Files Changed

• neuralmind/synapse_dynamics.py — NEW (920 lines, 6 techniques)
• neuralmind/retrieval_enhancement.py — NEW (600 lines, 5 fixes)
• neuralmind/context_selector.py — +70 lines (integration)

• neuralmind/core.py

  neuralmind/__init__.py — +10 lines (exports)

• tests/test_synapse_dynamics.py — NEW (26 tests)
• tests/test_retrieval_enhancement.py — NEW (28 tests)