NWcortex 1.0.4

NeuroWeave Cortex (NWC) — Graph-first cognitive memory runtime for AI agents. Hippocampal-inspired architecture with domain routing, spreading activation, edge budget management, 4-layer compression, thermal storage, personality modeling, and 8-phase sleep consolidation.

NeuroWeave Cortex (NWC) — Cognitive Memory Runtime

An infrastructure-grade memory system for AI agents. Not a vector database. Not a graph database. A memory runtime — it remembers, forgets, strengthens, connects, abstracts, and evolves, the way biological memory does.

Anchors: 131   Edges: 211   Ghosts: 0   Schemas: 9   Abstracts: 1
Memory Stability: 0.72   Recall Plasticity: 0.58   Compression: 1.6x

What makes it different

Vector databases retrieve. Graph databases traverse. NeuroWeave Cortex runs a cognitive lifecycle:

Capability	Vector DB	Graph DB	NeuroWeave Cortex
Semantic retrieval	yes	no	yes
Graph traversal	no	yes	yes
Automatic forgetting (survival decay)	no	no	yes
Memory strengthening (rehearsal)	no	no	yes
Conflict detection (contradiction edges)	no	no	yes
Multi-strategy RRF retrieval fusion	no	no	yes
Cross-Encoder reranking	no	no	yes
Explainable reasoning paths	no	no	yes
Memory revision engine (sleep)	no	no	yes
Hot/Warm/Cold memory tiering	no	no	yes
Multimodal (text+image+audio)	no	no	yes
Fuzzy recall ("I think I remember...")	no	no	yes
Emergent abstraction (pattern discovery)	no	no	yes
Temporal context (TimeSpine-indexed)	no	no	yes
8-phase sleep consolidation	no	no	yes
Ghost revival (savings effect)	no	no	yes
Autobiographical self-model	no	no	yes

Quick Start

from star_graph import MemoryManager
from star_graph.scheduler import AgentContext

# One-line setup
mgr = MemoryManager()

# Remember things
mgr.remember("User debugs Redis connection timeout — pool was 10, fixed to 20",
             tags=["redis", "debug", "timeout"])
mgr.remember("User knows Python async programming with asyncio",
             tags=["python", "knowledge"])
mgr.remember("User prefers type hints and concise code",
             tags=["preference", "style"])

# Working memory — fast, ephemeral buffer for active context
mgr.remember_working("Currently debugging auth middleware timeout",
                     tags=["debug", "auth"])

# Context-aware recall — working memory gets retrieval priority
ctx = AgentContext(task_type="debugging", active_goals=["fix Redis connection"])
memories = mgr.recall("Redis connection pool config", context=ctx)
print(memories.memory_summary)

# System-2 deep recall for exhaustive or low-confidence queries
memories = mgr.dual_recall("list all Redis-related issues", context=ctx)

# Micro-consolidation — incremental, non-blocking
mgr.micro_consolidate()

# Let the system sleep — 8-phase consolidation
report = mgr.sleep()
print(report)

# Persist
mgr.save("agent_memory.json")
mgr.load("agent_memory.json")

# Export to editable Markdown (GBrain-aligned)
mgr.export_markdown("memories/")

Architecture

Three-layer design. Each layer depends only on the one below.

Layer 3: Behavior    │  Cortex routing, memory gating, working memory,
    (cortex.py,      │  dual-channel retrieval, adaptive replay, autobiographical memory
     router.py,      │  "What should I recall right now, at what detail level?"
     gate.py,        │
     working_memory.py,│
     scheduler.py,   │
     autobiography.py)│
                     │
Layer 2: Cognitive   │  Hub abstraction, cascade recall, TimeSpine temporal index,
    (retriever.py,   │  sleep consolidation, evolution, ghost revival,
     sleep.py,       │  abstraction, community detection, competition,
     evolution.py,   │  conflict detection, memory revision, cross-encoder
     ghost.py,       │  "How do memories connect, strengthen, and fade?"
     abstraction.py, │
     community.py,   │
     competition.py, │
     conflict_detection.py,│
     memory_revision.py,   │
     cross_encoder.py,     │
     timespine.py,   │
     cascade.py,     │
     hub.py)         │
                     │
Layer 1: Storage     │  CRUD, persistence, ANN indexing, tiered storage,
    (graph.py,       │  BM25 keyword index, multi-level caching
     anchor.py,      │  "Where is this memory stored?"
     storage.py,     │
     sqlite_storage.py,│
     index.py,       │
     bm25.py,        │
     cognitive_cache.py,│
     tier.py)        │

Core modules

Module	Role
manager.py	High-level facade — remember(), recall(), sleep(), save()
runtime.py	Dependency container — manages all subsystem lifecycles
retrieval_pipeline.py	5-layer dimensional descent (L0→L4) with automatic degradation
scheduler.py	Context-aware retrieval with memory type selection
working_memory.py	Short-term buffer (15-item, 1h TTL) — auto-promotes to long-term
sleep.py	10-phase sleep: N1_Replay → N2_Merge → N2b_ConflictDetection → N2c_MemoryRevision → N3_Compression → N3d_SleepRebuild → REM_Emotion → N4_Prune (v1.0.4)
evolution.py	Survival-based decay (Ebbinghaus/Power-law/Exponential), belief transitions, interference
retriever.py	HybridFusion + OscillationResonance + VectorSimilarity + Personalized PageRank + explainable scores
dual_channel.py	System-1 (fast) + System-2 (deep) dual-channel retrieval with auto-trigger
bm25.py	Sparse keyword retrieval (BM25) with weighted reciprocal rank fusion for hybrid search
cross_encoder.py	Post-retrieval Cross-Encoder reranking (sentence-transformers) for precision improvement
conflict_detection.py	Semantic contradiction detection with overwrite/coexist/deprecate resolution strategies
memory_revision.py	Sleep-phase memory revision engine — surprise-prioritized, template/LLM re-summarization
ghost.py	Latent memory traces with fuzzy recall and contradiction tracking (NegativeGhost)
abstraction.py	Emergent category discovery from anchor clusters
community.py	Louvain community detection with centroid routing
anchor.py	Memory unit with 6-state lifecycle, 10-dim AnchorVector, multiplicative retention
graph.py	Star graph with RichEdge (temporal, causal, state-transition), Schema, ReflectionNode
timespine.py	Temporal index for O(days×buckets) time-scoped retrieval
cascade.py	Causal chain traversal across connected memory sequences
hub.py	Hierarchical hub-and-spoke abstraction (leaf→domain→global)
cortex.py	Partitioned memory cortices with independent sleep and retrieval
cognitive_cache.py	Multi-level cache (query/session/topic/activation) + exact-match entity lookup
tier.py	STM/MTM/LTM/Core cognitive tiering + HOT/WARM/COLD storage tiers
autobiography.py	Self-narrative formation and autobiographical memory
atom_facts.py	LLM-powered atomic fact extraction from memory clusters
survival.py	Pluggable survival functions (Ebbinghaus, Power-law, Exponential, Custom)
compression.py	Multi-level session compression (episodic/strategic/meta)
resonance.py	Phase-locked oscillation resonance for temporal-coherent retrieval
streaming.py	Streaming memory buffer with backpressure
benchmark.py	Built-in benchmark suite (5 categories)
batch_vectorizer.py	Deferred batch embedding writes (buffer >= 32 / >30s flush) with SQLite WAL crash recovery
zero_llm_pipeline.py	7-stage zero-LLM ingestion: security → embed → dedup → entity → classify → score → link
multimodal.py	Cross-modal text/image/audio memory — CLIP image encoding, Whisper/spectrogram audio, unified embedding space
forget_certificate.py	Ed25519-signed JWS deletion certificates — GDPR Article 17 provable erasure
markdown_export.py	Operator-owned plain-text memory export (GBrain-aligned); timeline + topic Markdown
config.py	Centralized YAML config with schema validation, dot-path access, overrides

Retrieval Benchmarks

LoCoMo Benchmark (real-world conversations)

Evaluated on the LoCoMo-10 dataset: 10 conversations, 5,882 turns across 272 sessions, 1,986 QA pairs across 5 categories. Pure retrieval (no LLM generation).

Method	has_answer	F1	Latency
VectorSimilarity	13.1%	0.026	122.2ms
OscillationResonance	11.9%	0.026	110.4ms
HybridFusion + BM25	15.3%	0.025	<200ms

Per-category has_answer:

Category	#QA	VecSim	OscRes	HybFus
Temporal (1)	282	3.5%	2.8%	4.3%
Short Memory (2)	321	1.9%	2.2%	1.9%
Long Memory (3)	96	2.1%	2.1%	3.1%
Composite (4)	841	18.0%	16.4%	21.0%
Adversarial (5)	446	20.4%	18.4%	23.5%

Internal Benchmark (synthetic multi-session)

6 sessions × 80 turns, 5 categories. 1.7x compression (6,708 → 3,982 tokens) with maintained or improved recall.

Method	C-R@3	C-R@5	Interf
VectorSimilarity	0.933	0.933	N/A
OscillationResonance	0.967	0.967	0.667
HybridFusion	0.900	0.900	0.667

Memory Lifecycle

Every anchor moves through 6 states:

ACTIVE → REHEARSING → CONSOLIDATING → DORMANT → GHOST → REACTIVATED

• Active: Just created or recently accessed — fully plastic, easy to update
• Rehearsing: Being replayed during sleep — temporarily elevated importance
• Consolidating: Transferring from hippocampal to cortical — increasing stability
• Dormant: Stable, low-activity — read-only, cortical retrieval
• Ghost: Pruned but with residual trace — can partially recall or fully revive
• Reactivated: Ghost revived by new relevant experience — reduced stability, high plasticity

Paired with ThermalState (HOT → WARM → COLD → DEAD) for storage tier switching:

• HOT: in-memory, fully accessible
• WARM: in-memory, periodically flushed to disk
• COLD: disk-only, transparent thaw on access

Sleep Consolidation

Sleep is not cleanup. Sleep changes the graph:

1. N1_Replay — prioritizes surprising and emotional memories for replay via SWR scoring
2. N2_Merge — fuses near-duplicate anchors (ANN-accelerated, O(n×k)), bridges constellations
3. N3_Compression — transfers memories from hippocampal to cortical, forms schemas
4. N3b_AtomFacts — LLM extraction of atomic facts from compressed clusters
5. REM_Emotion — strips emotional charge from consolidated memories
6. N4_Prune — removes weak anchors/edges, creates ghost traces for savings effect
7. N5_HubConnect — cross-cortex hub bridge formation
8. N6_IndexRebuild — refreshes ANN, BM25, and community indices

Dual-Channel Retrieval

System-1 (fast, embedding + BM25 hybrid) and System-2 (deep, hierarchical traversal) with automatic triggering:

• Low-confidence System-1 results (<0.35) automatically trigger System-2
• Structural keywords ("all", "every", "list", "which", "before", "last") trigger exhaustive search
• Results merged via weighted reciprocal rank fusion

Configuration

from star_graph.config import config, override, load_config

# Dot-path access
threshold = config.sleep.merge.default_threshold  # 0.85

# Programmatic override
override('sleep.merge.default_threshold', 0.75)
override('gate.k', 30)

# Schema validation
warnings = config.validate()  # type, range, and cross-section checks

# Load custom YAML
cfg = load_config("my_params.yaml")

See star_graph/defaults.yaml for all 300+ tunable parameters.

Installation

# From PyPI (package name: NWcortex, import as: star_graph)
pip install NWcortex

# With sentence-transformers for semantic embeddings
pip install "NWcortex[embeddings]"

# With MCP server support (for AI agent integration)
pip install "NWcortex[mcp]"

# Everything
pip install "NWcortex[all]"

# Run demo
python examples/emergence_demo.py

Note: Without [embeddings], the system uses a lightweight TF-IDF fallback for text encoding. Install sentence-transformers only if you need semantic-quality embeddings.

MCP Server — AI Agent Integration

NeuroWeave Cortex ships with a Model Context Protocol server. Connect any MCP-compatible AI agent (Claude Desktop, OpenClaw, Cursor, etc.) for persistent cognitive memory across conversations.

Quick start

# Install with MCP support
pip install "NWcortex[mcp]"

# Launch the MCP server on stdio
nwc-mcp

# With persistent storage
nwc-mcp --storage agent_memory.json

# Load a previously saved memory graph
nwc-mcp --load my_memories.json

Tool reference (12 tools)

Tool	Description
remember	Store a long-term memory (tags, importance, emotional valence)
remember_working	Store in fast ephemeral working-memory buffer (current task context)
recall	Context-aware semantic retrieval with task-type routing
forget	Remove a memory, creating a ghost trace for potential fuzzy recall
sleep	Run 5-phase sleep consolidation — merges, prunes, forms schemas
consolidate	Micro-consolidation — incremental, non-blocking
stats	Memory system statistics (anchors, edges, ghosts, schemas, cognitive health)
fuzzy_recall	Low-confidence recall from ghost traces ("I seem to remember...")
get_profile	Inferred user profile from accumulated memories
evolve	Memory evolution cycle (decay, boost, conflict resolution)
save	Persist memory graph to disk (JSON)
load	Load memory graph from disk (JSON)

Claude Desktop config

{
  "mcpServers": {
    "neuroweave-cortex": {
      "command": "nwc-mcp",
      "args": ["--storage", "/path/to/agent_memory.json"]
    }
  }
}

OpenClaw config

{
  "mcpServers": {
    "neuroweave-cortex": {
      "command": "nwc-mcp",
      "args": ["--storage", "/path/to/agent_memory.json"]
    }
  }
}

REST API Server

# Start the REST server
nwc-server --port 8420

# Or via module
python -m star_graph.server --port 8420

Endpoints:

Method	Path	Description
GET	/health	Health check with anchor/edge counts
GET	/metrics	Prometheus-format metrics
GET	/stats	Full memory system statistics
POST	/remember	Store a memory {"text": "...", "tags": [...]}
POST	/recall	Retrieve memories {"query": "...", "max_items": 10}
POST	/sleep	Run sleep consolidation
POST	/consolidate	Run micro-consolidation

Interactive Demo

python examples/emergence_demo.py

# Or use the CLI:
# sg-add "Discussed microservices deployment patterns" --tags architecture --emotional 0.6
# sg-query "database connection pooling best practices"
# sg-query --trace "When did Alice visit Hawaii?"
# sg-stats --schemas --ghosts
# sg-sleep --retention 0.15 --edge-prune 0.1
# nwc-export --output memories/ --organize both
# nwc-forget <anchor_id> --certificate --reason gdpr_art17
# nwc-verify certificates/forget-<id>.jws

Benchmarks

python examples/memory_benchmark.py --quick   # 4 sessions, ~200 turns
python examples/memory_benchmark.py --full    # 12 sessions, ~5000 turns

Running Tests

pip install pytest pytest-cov
pytest tests/ -v

# With coverage report
pytest tests/ --cov=star_graph --cov-report=term

Status: 1,989 tests passing, 84 modules (v1.0.4-dev).

Roadmap

See ROADMAP.md for planned work.

License

MIT