entroly 0.3.0

Information-theoretic context optimization for AI coding agents. Knapsack-optimal token budgeting, Shannon entropy scoring, SimHash dedup, predictive pre-fetch. MCP server.

Entroly

Information-theoretic context optimization for AI coding agents.

Every AI coding tool manages context with dumb FIFO truncation — stuffing tokens until the window is full, then cutting from the top. Entroly applies mathematics to select the optimal context subset.

pip install entroly

Architecture

Hybrid Rust + Python: CPU-intensive math (knapsack DP, entropy, SimHash, LSH, dependency graph) runs in Rust via PyO3 for 50-100x speedup. MCP protocol and orchestration run in Python via FastMCP. Pure Python fallbacks activate automatically if the Rust extension isn't available.

What It Does

An MCP server that sits between your AI coding tool and the LLM, providing:

Engine	What it does	How it works
Knapsack Optimizer	Selects mathematically optimal context subset	0/1 Knapsack DP with budget quantization (N ≤ 2000), greedy fallback (N > 2000)
Entropy Scorer	Measures information density per fragment	Shannon entropy (40%) + boilerplate detection (30%) + cross-fragment multi-scale n-gram redundancy (30%)
SimHash Dedup	Catches near-duplicate content in O(1)	64-bit SimHash fingerprints with 4-band LSH bucketing, Hamming threshold = 3
Multi-Probe LSH Index	Sub-linear semantic recall over 100K+ fragments	12-table LSH with 10-bit sampling + 3-neighbor multi-probe queries
Dependency Graph	Pulls in related code fragments together	Symbol table + auto-linking (imports, type refs, function calls) + two-pass knapsack refinement
Predictive Pre-fetch	Pre-loads context before the agent asks	Static import analysis + test file inference + learned co-access patterns
Checkpoint & Resume	Crash recovery for multi-step tasks	Gzipped JSON state serialization (~100 KB per checkpoint)
Feedback Loop	Learns which context leads to good outputs	Wilson score lower-bound confidence intervals (same formula as Reddit ranking)
Context Ordering	Orders fragments for optimal LLM attention	Pinned → criticality level → dependency count → relevance score
Guardrails	Auto-pins safety-critical files, classifies tasks	Criticality levels (Safety/Critical/Important/Normal) + task-aware budget multipliers
PRISM Optimizer	Adapts scoring weights to the codebase	Anisotropic spectral optimization via Jacobi eigendecomposition on 4×4 covariance matrix
Provenance Chain	Detects hallucination risk in selected context	Tracks source verification + confidence scoring per fragment

Setup

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "entroly": {
      "command": "entroly"
    }
  }
}

Claude Code

claude mcp add entroly -- entroly

Cline / Any MCP Client

{
  "entroly": {
    "command": "entroly",
    "args": []
  }
}

MCP Tools

remember_fragment

Store context with auto-dedup, entropy scoring, dependency linking, and criticality detection.

remember_fragment(content="def process_payment(...)...", source="payments.py", token_count=45)
→ {"status": "ingested", "entropy_score": 0.82}

remember_fragment(content="def process_payment(...)...")  # same content
→ {"status": "duplicate", "duplicate_of": "a1b2c3", "tokens_saved": 45}

optimize_context

Select the optimal context subset for a token budget. Includes dependency boosting, ε-greedy exploration, context sufficiency scoring, and provenance metadata.

optimize_context(token_budget=128000, query="fix payment bug")
→ {
    "selected_fragments": [...],
    "optimization_stats": {"method": "exact_dp", "budget_utilization": 0.73},
    "tokens_saved_this_call": 42000,
    "sufficiency": 0.91,
    "hallucination_risk": "low"
  }

recall_relevant

Sub-linear semantic recall via multi-probe LSH. Falls back to brute-force scan on cold start.

recall_relevant(query="database connection pooling", top_k=5)
→ [{"fragment_id": "...", "relevance": 0.87, "content": "..."}]

record_outcome

Feed the Wilson score feedback loop. Adjusts fragment scoring multipliers in the range [0.5, 2.0].

record_outcome(fragment_ids=["a1b2c3", "d4e5f6"], success=true)
→ {"status": "recorded", "fragments_updated": 2}

explain_context

Per-fragment scoring breakdown with sufficiency analysis and exploration swap log.

explain_context()
→ {
    "fragments": [{"id": "...", "recency": 0.9, "frequency": 0.3, "semantic": 0.7, "entropy": 0.8}],
    "sufficiency": 0.91,
    "exploration_swaps": 1
  }

checkpoint_state / resume_state

Save and restore full session state — fragments, dedup index, co-access patterns, feedback scores.

checkpoint_state(task_description="Refactoring auth module", current_step="Step 5/8")
→ {"status": "checkpoint_saved", "fragments_saved": 47}

resume_state()
→ {"status": "resumed", "restored_fragments": 47, "metadata": {"step": "Step 5/8"}}

prefetch_related

Predict and pre-load likely-needed context using import analysis, test file inference, and co-access history.

prefetch_related(file_path="src/payments.py", source_content="from utils import...")
→ [{"path": "src/utils.py", "reason": "import", "confidence": 0.70}]

get_stats

Session statistics and cost savings.

get_stats()
→ {
    "fragments": 142,
    "total_tokens": 384000,
    "savings": {
      "total_tokens_saved": 284000,
      "total_duplicates_caught": 12,
      "estimated_cost_saved_usd": 0.85
    }
  }

The Math

Multi-Dimensional Relevance Scoring

Each fragment is scored across four dimensions:

r(f) = (w_rec · recency + w_freq · frequency + w_sem · semantic + w_ent · entropy)
       / (w_rec + w_freq + w_sem + w_ent)
       × feedback_multiplier

Default weights: recency 0.30, frequency 0.25, semantic 0.25, entropy 0.20.

• Recency: Ebbinghaus forgetting curve — exp(-ln(2) × Δt / half_life), half_life = 15 turns
• Frequency: Normalized access count (spaced repetition boost)
• Semantic similarity: SimHash Hamming distance to query, normalized to [0, 1]
• Information density: Shannon entropy + boilerplate + redundancy (see below)

Knapsack Context Selection

Context selection is the 0/1 Knapsack Problem:

Maximize:   Σ r(fᵢ) · x(fᵢ)     for selected fragments
Subject to: Σ c(fᵢ) · x(fᵢ) ≤ B  (token budget)

Two strategies based on fragment count:

• N ≤ 2000: Exact DP with budget quantization into 1000 bins — O(N × 1000)
• N > 2000: Greedy density sort — O(N log N), Dantzig 0.5-optimality guarantee

Pinned fragments (safety-critical files, config files) are always included; remaining budget is allocated via DP/greedy.

Shannon Entropy Scoring

Three components combined:

score = 0.40 × normalized_entropy + 0.30 × (1 - boilerplate_ratio) + 0.30 × (1 - redundancy)

• Shannon entropy (40%): H = -Σ p(char) · log₂(p(char)), normalized by 6.0 bits/char. Stack-allocated 256-byte histogram, single O(n) pass.
• Boilerplate detection (30%): Pattern matching for imports, pass, dunder methods, closing delimiters.
• Cross-fragment redundancy (30%): Multi-scale n-gram overlap with adaptive weights by fragment length — bigram-heavy for short fragments (<20 words), 4-gram-heavy for long fragments (>100 words). Parallelized with rayon.

SimHash Deduplication

64-bit fingerprints from word trigrams hashed via MD5:

• Hamming distance ≤ 3 → near-duplicate
• 4-band LSH bucketing for O(1) candidate lookup
• Separate 12-table multi-probe LSH index for semantic recall (~3 μs over 100K fragments)

Dependency Graph

Auto-linking via source analysis:

• Imports (strength 1.0): Python from X import Y, Rust use, JS import
• Type references (0.9): Type annotations, isinstance checks
• Function calls (0.7): General identifier usage matching against symbol table
• Same module (0.5): Co-located definitions

Two-pass knapsack refinement: initial selection → boost dependencies of selected fragments → re-optimize.

Task-Aware Budget Multipliers

Bug tracing / debugging     → 1.5× budget
Exploration / understanding → 1.3× budget
Refactoring / code review   → 1.0× budget
Testing                     → 0.8× budget
Code generation             → 0.7× budget
Documentation               → 0.6× budget

PRISM Spectral Optimizer

Tracks a 4×4 covariance matrix over scoring dimensions [recency, frequency, semantic, entropy] with EMA updates (β=0.95). Jacobi eigendecomposition finds principal axes. Anisotropic spectral gain dampens noisy dimensions and amplifies clean signals — automatic learning rate adaptation without hyperparameter tuning.

Configuration

EntrolyConfig(
    default_token_budget=128_000,     # GPT-4 Turbo equivalent
    max_fragments=10_000,             # session fragment cap
    weight_recency=0.30,              # scoring weights (sum to 1.0)
    weight_frequency=0.25,
    weight_semantic_sim=0.25,
    weight_entropy=0.20,
    decay_half_life_turns=15,         # Ebbinghaus half-life
    min_relevance_threshold=0.05,     # auto-evict below this
    dedup_similarity_threshold=0.92,
    prefetch_depth=2,
    max_prefetch_fragments=10,
    auto_checkpoint_interval=5,       # checkpoint every N tool calls
)

References

• Shannon (1948) — Information Theory
• Charikar (2002) — SimHash
• Ebbinghaus (1885) — Forgetting Curve
• Dantzig (1957) — Greedy Knapsack Approximation
• Wilson (1927) — Score Confidence Intervals
• ICPC (arXiv 2025) — In-context Prompt Compression
• Proximity (arXiv 2026) — LSH-bucketed Semantic Caching
• RCC (ICLR 2025) — Recurrent Context Compression
• ILRe (ICLR 2026) — Intermediate Layer Retrieval
• Agentic Plan Caching (arXiv 2025)

Part of the Ebbiforge Ecosystem

Entroly integrates with hippocampus-sharp-memory for persistent memory and Ebbiforge for TF embeddings and RL weight learning. Both are optional — Entroly works standalone with pure Python fallbacks.

License

MIT