bilinc 1.1.0

Verifiable State Plane for Autonomous Agents

Bilinc

Verifiable state plane for autonomous agents

Bilinc manages truth state for long-running AI agents. Not just memory — verified, auditable, contradiction-resistant state.

pip install bilinc

---

What Bilinc Does

Most agent memory systems are vector search with a wrapper. Bilinc is different:

• Z3 formal verification on every state commit (catches contradictions before they're stored)
• AGM belief revision (EXPAND/CONTRACT/REVISE) resolves conflicts without silent data loss
• Hybrid decay model (exponential → power-law) with LTP protection for important memories
• Knowledge graph with spreading activation for multi-hop retrieval
• FTS5 + vector search hybrid recall with RRF fusion
• Blind spot detection finds knowledge gaps before they matter
• Rollback, snapshot, diff — full state history recovery
• MCP server over stdio and authenticated HTTP

Benchmark Results

Benchmark	Score	Details
LongMemEval	98.0%	No LLM, fully local
ConvoMem	98.0%	5 categories, real recall pipeline
LoCoMo	90.3%	Temporal + causal + multi-hop

Score progression: ConvoMem 17.5% → 98.0%, LoCoMo 9.1% → 90.3%

Quick Start

from bilinc import StatePlane

# Create a state plane
sp = StatePlane(backend="sqlite", db_path="agent_state.db")

# Commit a belief (auto-verified)
sp.commit("user_preference", {
    "theme": "dark",
    "language": "python"
}, memory_type="semantic", importance=0.9)

# Recall with hybrid search (FTS5 + vector + KG)
results = sp.recall("user preference", limit=5)

# Check for contradictions
from bilinc import verify
report = verify.check_consistency(sp)

# Snapshot and rollback
snapshot = sp.snapshot()
sp.rollback(snapshot)

Architecture

StatePlane (orchestrator)
├── WorkingMemory (8 slots, PFC-inspired)
├── SQLite Backend (persistent, FTS5, sqlite-vec)
├── AGM Belief Revision (EXPAND/CONTRACT/REVISE)
├── Knowledge Graph (entities + relations)
├── Verification Gate (Z3 SMT solver)
├── Audit Trail (full provenance)
└── MCP Server (stdio + HTTP)

Memory Types (5-type taxonomy)

Type	Decay	Use Case
Working	24h half-life	Active context buffer
Episodic	7 days	Conversation events
Procedural	30 days	How-to knowledge
Semantic	180 days	Facts and entities
Spatial	90 days	Location/environment

Hybrid Recall (3-level)

Query → Level 1: FTS5 keyword (BM25 + porter stemming)
      → Level 2: Vector similarity (sqlite-vec + Ollama)
      → Level 3: KG spreading activation (HippoRAG-inspired)
      → RRF fusion → Decay-aware reranking → Results

MCP Integration

Bilinc ships as an MCP server for Claude Code, Cursor, and any MCP-compatible agent:

{
  "mcpServers": {
    "bilinc": {
      "command": "python",
      "args": ["-m", "bilinc.mcp_server.server_v2"],
      "env": {"BILINC_DB_PATH": "~/bilinc.db"}
    }
  }
}

11 MCP tools: commit_mem, recall, revise, forget, consolidate, contradictions, diff, snapshot, status, verify, query_graph

Modules (v1.1.0)

Module	Purpose
core/decay.py	Hybrid decay model + LTP protection
core/blind_spots.py	Knowledge gap detection
core/consolidation.py	3-stage consolidation pipeline
core/vector_search.py	Hybrid search (FTS5 + vector + RRF)
core/kg_retrieval.py	KG spreading activation
core/temporal.py	Temporal reasoning (Allen's interval algebra)
storage/sqlite.py	SQLite + FTS5 + sqlite-vec
storage/postgres.py	PostgreSQL backend

Testing

pip install -e ".[dev]"
pytest tests/ -v

217 tests passing, 0 failures.

What Makes Bilinc Different

Feature	Bilinc	Mem0	Zep	Letta
Z3 formal verification	✅	❌	❌	❌
AGM belief revision	✅	❌	❌	❌
Rollback/snapshot/diff	✅	❌	❌	❌
Audit trail	✅	❌	❌	❌
Hybrid decay (exp→power-law)	✅	❌	❌	❌
Blind spot detection	✅	❌	❌	❌
FTS5 + vector hybrid	✅	❌	✅	❌
Knowledge graph	✅	❌	✅	❌
MCP server	✅	❌	❌	✅
Fully local (no cloud)	✅	❌	❌	✅

Others store memories. Bilinc manages truth.

License

MIT