SOMA v2.0.14 — The Cognitive Kernel for AI Agents

Wisdom over Memory — 智慧超越记忆
Five capability lines. Seven thinking laws. One cognitive kernel that remembers, reasons, collaborates, and evolves.

pip install soma-wisdom     # 5 minutes from zero to thinking agent

👆 点击即可在浏览器中体验 SOMA 面板 — 无需安装，一分钟启动

from soma import SOMA

soma = SOMA()
soma.remember("First-principles thinking: deconstruct to fundamentals...",
              context={"domain": "philosophy"}, importance=0.9)
answer = soma.respond("How to analyze our growth bottleneck?")
# → decomposes through 7 thinking laws → activates relevant memories → returns reasoned answer

Why SOMA instead of a vector database? Traditional memory (ChromaDB, Mem0) stores and retrieves. SOMA thinks first: a 7-law reasoning network decomposes problems before fetching memories. The result: agents that systematically analyze, not just pattern-match.

📖 中文文档 | Docs | Demo | Roadmap | Changelog | Contributing

v2.0.6 — Code Memory + Active Memory Management

v2.0.6 adds AST-aware code memory, Ebbinghaus forgetting curves, and cross-agent CLI on top of v2.0.5's autonomous cognitive loop:

• Code Memory: AST parsing → structured code memory with semantic triples (CALLS, INHERITS, HAS_METHOD)
• Memory Lifecycle: Forgetting curves (7/30/90-day half-life), episodic→semantic consolidation, conflict detection, auto-pruning
• Unified CLI: soma recall/record/think/maintain — all 5 dev agents share one wisdom network
• Health Monitoring: Real-time ONNX latency tracking (is_healthy, latency_report())

853 tests passed. Semantic recall 10/10 perfect. The autonomous cognitive loop completes full reasoning without external LLM calls:

Benchmark (v2.0.5 on 17K+ memories, 748 tests): Overall 87.5, Wisdom 85.6 (+1.9 counterfactual), Semantic Recall 10/10.

Every capability line that started as a seed in v0.1 has grown into a complete system:

All new features default to off. Upgrade from any 0.x version with zero code changes.

Architecture

┌──────────────────────────────────────────────────────────────────────────────┐
│                         SOMA v2.0.14 — Cognitive Kernel                         │
│                                                                                │
│  ┌──────────────────────────────────────────────────────────────────┐        │
│  │  L3 User Profile — "Knows who you are"                             │        │
│  │  Auto-extracted traits: preferences, skills, knowledge gaps, goals │        │
│  └──────────────────────────────────────────────────────────────────┘        │
│                                    ↑                                          │
│  ┌──────────────────────────────────────────────────────────────────┐        │
│  │  L2 Scene Blocks — "Understands your context"                      │        │
│  │  Auto-aggregated work contexts ("building a Python data platform") │        │
│  └──────────────────────────────────────────────────────────────────┘        │
│                                    ↑                                          │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────────────────────┐    │
│  │ WisdomEngine  │→│ActivationHub │→│  L1 Episodic Memory (the original) │    │
│  │ · 7 thinking  │  │ · bidirectional│ · SQLite + vector + FTS5           │    │
│  │   laws        │  │   activation  │ · weighted RRF + time decay         │    │
│  │ · law chaining│  │ · conflict    │ · knowledge graph + causal chains   │    │
│  │ · combo synth │  │   detection   │ · cross-domain analogy engine       │    │
│  │ · complexity  │  │ · MMR re-rank │                                      │    │
│  │   assessment  │  │ · frame nudge │                                      │    │
│  └──────────────┘  └──────────────┘  └──────────────────────────────────┘    │
│         │                  │                        │                        │
│         ▼                  ▼                        ▼                        │
│  ┌──────────────────────────────────────────────────────────────────┐       │
│  │  Multi-Agent Layer          │  Evolution Loop                     │       │
│  │  · AgentRegistry            │  · MetaEvolver: bias detect →       │       │
│  │  · ExpertRouter (3-tier)    │    re-weight → solidify → share     │       │
│  │  · ConsensusProtocol        │  · CapturePipeline: auto L2+L3      │       │
│  │  · DistributedEvolver       │  · FrameAnchoringDetector           │       │
│  └──────────────────────────────────────────────────────────────────┘       │
└──────────────────────────────────────────────────────────────────────────────┘

Fifteen-Stage Wisdom Pipeline:
  Assess → Decompose → Chain → Combine → Semantic-fallback
         → Context-sort → Activate → Conflict-detect → Frame-nudge → Anti-bias
         → Reason → Synthesize → Causal-extract → Backward-propagate → Evolve

Screenshots

🧠 Wisdom Chat  ·  📊 5D Benchmark  ·  💻 IDE Integration  ·  🔌 REST API

Wisdom Chat — 7 thinking laws decompose problems, bidirectional memory activation, LLM streaming	5D Benchmark — Memory/Wisdom/Evolution/Scalability/Overall, live competitor comparison
IDE Integration — Claude Code / VS Code one-click access, auto-persistent memory	REST API — FastAPI + SSE streaming, multi-model management, API Key auth

Dashboard

Start the API server and open http://localhost:8765:

SOMA_API_KEY=test python dash/server.py

Vue 3 dashboard with i18n (English / 中文), 6 views: Chat · Framework · Memory · Analytics · Benchmark · Settings.

Installation

pip install soma-wisdom

Requires Python 3.10+. The embedding engine uses ONNX Runtime for CPU inference — no CUDA, no Docker, no external services.

First run downloads a small ONNX model (~100 MB, Chinese-English bilingual).

python -m soma          # verify everything works in one command
soma-quickstart         # or use the CLI entry point

Core Concepts

1. Wisdom Framework — 7 Thinking Laws

Customize in wisdom_laws.yaml (bundled in the package — always available).

Thinking laws form a reasoning network. When a law triggers, its relations propagate activation to related laws (×0.35–0.50 bonus). When two laws fire together, synthesized perspectives emerge (e.g., "First Principles × Systems Thinking → Root-Cause System Analysis"). Weights auto-tune based on actual success/failure — a startup team and a large enterprise will naturally evolve different weight distributions.

2. Three-Tier Memory System

v1.0 introduces automatic memory layering — the system quietly builds understanding while you work:

The entire process is automatic. You use SOMA normally; behind the scenes, CapturePipeline aggregates fragments into scenes and distills scenes into your profile.

3. Bidirectional Activation — Hybrid RRF

Memories are matched through weighted Reciprocal Rank Fusion:

• Vector semantic similarity (×2 weight) via ONNX embeddings
• Keyword exact match (×1 weight)
• Knowledge graph expansion (×0.5 weight)

All three paths compete and complement, producing scores that reflect true relevance — not just keyword overlap.

4. Meta-Evolution — Self-Optimization

SOMA tracks success/failure of each thinking law across sessions. Every 5 sessions, evolve() automatically:

• Memory consolidation: similar memories auto-merged, reducing redundancy
• Active forgetting: low-value memories archived with Ebbinghaus decay curves
• Bias detection: laws used >40% of the time get penalized (-0.05) to prevent thinking ruts; underused high-success laws get boosted (+0.03)
• Dynamic step sizing: adjustment magnitude scales with sample count (0.01 → 0.02 → 0.03)
• Skill solidification: successful (law, domain, outcome) patterns become permanent skills after 3+ occurrences

5. Knowledge Graph & Reasoning Engine

Six cognitive capabilities that upgrade SOMA from a memory store to a reasoning system:

• Causal Chain Tracing: Follow the causal graph to root causes — "why did user churn rise?" → traces back to "an API change three months ago slowed response times"
• Conflict Detection: When memories contain contradictory claims, flag and down-weight automatically, preventing the LLM from being misled
• Cross-Domain Analogy: Map structural patterns across domains — e.g., "supply chain bottleneck" ≈ "blood vessel blockage"
• Quality Evaluation: Self-score each reasoning output on consistency, coherence, and actionability. Below-threshold outputs trigger reflective improvement
• Graph Retrieval Expansion: Keywords are no longer isolated — BFS traversal (depth=2) discovers neighbor concepts, breaking retrieval silos
• Backward Propagation: High-activation memories suggest new thinking foci — the memory→focus feedback loop discovers perspectives the initial decomposition missed

6. Multi-Agent Collaboration

Multiple SOMA agents work as a team:

• Expert Specialization: Each agent has independent memory, independent evolution path, and domain expertise
• Automatic Routing: Problems dispatched to the most suitable expert via keyword + semantic matching — zero LLM involvement, completed in milliseconds
• Consensus Protocols: When experts disagree — L1 weighted voting → L2 LLM arbitration → L3 dialectic synthesis
• Distributed Evolution: Each expert evolves independently while periodically merging global experience
• Memory Isolation: Three-state isolation via agent_id + group_id — private, group-shared, and global

7. Sunyata Awareness Layer

SOMA detects when you're over-anchored to a single cognitive frame and gently nudges — without blocking, forcing, or changing the pipeline.

• 8 cognitive frame pairs: Technical/Business, Management/Legal, Short-term/Long-term, Internal/External
• Pure keyword matching: Zero LLM/embedder dependency
• Low-interference: Blockquote footnote at prompt end — won't dominate the reasoning flow
• Default off: enable_frame_detection=False, 100% backward compatible

soma = SOMA()
soma._agent.config.enable_frame_detection = True  # opt-in
# SOMA now notices when you're stuck in one perspective

API Reference

SOMA Facade (Python SDK)

from soma import SOMA

soma = SOMA(
    framework_config=None,        # default: bundled wisdom_laws.yaml
    llm="deepseek-chat",          # LiteLLM model string
    use_vector_search=True,       # ONNX semantic search
    persist_dir="soma_data",      # persistence directory
    recall_threshold=0.01,        # minimum activation score
    top_k=5,                      # default recall count
    agent_id="",                  # v1.0: agent identity for multi-agent
    group_id="",                  # v1.0: group for shared memory
)

# Wisdom pipeline
soma.respond(problem: str) -> str
soma.chat(problem: str) -> dict          # structured: foci + memories + weights

# Memory operations
soma.remember(content, context, importance) -> str  # returns memory_id
soma.remember_semantic(subject, predicate, object_, confidence)
soma.query_memory(query: str, top_k: int) -> list

# v1.0: Three-tier memory
soma.get_scenes(user_id="", top_k=10) -> list
soma.get_profile(user_id="") -> list
soma.capture_scenes(user_id="", force=False) -> int
soma.update_profile(user_id="", force=False) -> int

# Introspection & evolution
soma.decompose(problem: str) -> list     # show thinking dimensions
soma.reflect(task_id, outcome) -> None   # record outcome for evolution
soma.evolve() -> list                    # trigger automatic weight adjustment
soma.get_weights() -> dict               # current law weights
soma.adjust_weight(law_id, new_weight)   # manual override
soma.discover_laws() -> dict | None      # autonomous law discovery
soma.approve_law(candidate) -> bool      # approve a discovered law
soma.stats -> dict                       # memory store statistics

REST API (Language-Agnostic)

# Start server
SOMA_API_KEY=your-key python dash/server.py    # → http://localhost:8765

# Standard chat
curl -X POST http://localhost:8765/api/chat \
  -H "X-API-Key: your-key" \
  -H "Content-Type: application/json" \
  -d '{"problem": "How to improve team productivity?"}'

# SSE streaming (decompose → activate → delta → done)
curl -X POST http://localhost:8765/api/chat/stream \
  -H "X-API-Key: your-key" \
  -H "Content-Type: application/json" \
  -d '{"problem": "Analyze our growth bottleneck"}'

# Memory search
curl -X POST http://localhost:8765/api/memory/search \
  -H "X-API-Key: your-key" \
  -H "Content-Type: application/json" \
  -d '{"query": "growth strategy", "top_k": 10}'

Set SOMA_API_KEY env var to enable authentication. Leave unset for local development.

LangChain Tool

from soma.langchain_tool import create_soma_tool
from soma.agent import SOMA_Agent
from soma.config import SOMAConfig, load_config

agent = SOMA_Agent(SOMAConfig(framework=load_config()))
tool = create_soma_tool(agent)
result = tool.run("Analyze this problem...")

AI Coding Agent Integration (Claude Code / VS Code / JetBrains)

SOMA runs alongside AI coding tools as a persistent wisdom backend — it learns from every debug session, code review, and architectural decision.

# 1. Start SOMA server (once)
SOMA_API_KEY=dev-key python dash/server.py

Claude Code (Agent SDK) — use as a custom MCP server or REST tool:

# In your custom Claude Agent tool — persist insights across sessions
import requests
requests.post("http://localhost:8765/api/memory/remember",
    headers={"X-API-Key": "dev-key"},
    json={"content": "Bug: N+1 query in OrderService.getOrders, root cause: missing @BatchSize on items relation, fix: add Hibernate batch annotation + integration test", "importance": 0.9})

Any IDE / CLI tool — just curl the REST API. No SDK required.

# Record a debug finding
curl -s -X POST http://localhost:8765/api/memory/remember \
  -H "X-API-Key: dev-key" -H "Content-Type: application/json" \
  -d '{"content":"Race condition in WebSocket handler: concurrent map writes on client.buf, root cause: missing mutex on write path, fix: sync.RWMutex around buffer ops","importance":0.9}'

# Recall relevant knowledge before starting a new task
curl -s -X POST http://localhost:8765/api/memory/search \
  -H "X-API-Key: dev-key" -H "Content-Type: application/json" \
  -d '{"query":"concurrency websocket golang","top_k":5}'

Real-World Impact

SOMA has been used in production across two distinct codebases — a Go-based CLI agent and a Python data platform — with the following results:

• Bug reoccurrence dropped significantly. When a bug fix is recorded as a SOMA memory (root cause + fix pattern), later sessions on the same codebase automatically retrieve it. Developers no longer re-debug the same class of issues across sprints.
• Architectural decisions became searchable. Each trade-off ("chose SQLite WAL over multi-worker for simplicity") is persisted with rationale context. Six months later, new team members can query "why single worker?" and get the original reasoning — not folklore.
• Cross-project knowledge transfer worked. A concurrency pattern learned in one codebase (the Go project) activated during debugging in the other (the Python project), because SOMA's semantic search matched "race condition" across language boundaries.
• Zero adoption friction. The Python project integrated via pip install soma-wisdom + 3 lines of code. The Go project integrated via REST API with a thin HTTP client (~50 lines). Neither required schema design, vector database setup, or infrastructure changes.
• Evolution is automatic. After ~50 sessions, SOMA's auto-weighting surfaced that "Inversion" (thinking backwards from failure) was consistently the most useful lens for debugging tasks, while "Analogical Reasoning" shined during architecture discussions. The framework self-tuned — no manual knob-twisting needed.

Benchmarks

SOMA v1.0 — benchmarked with 1,050 production memories from digital twin project testing (5 runs, statistical output):

Overall Score: 87.5/100

Key Metrics

Live Competitor Comparison

SOMA is the only system combining a reasoning framework, three-tier memory, causal analysis, conflict detection, cross-domain analogy, evolutionary self-optimization, multi-agent collaboration, and frame anchoring awareness — all without external services.

Full report: CHANGELOG.md

Reproducibility

pip install soma-wisdom chromadb
python -m soma.benchmarks --full --runs 5 --output reports/    # statistical benchmark
python scripts/live_benchmark.py --full --output reports/       # live competitor test

Development

git clone https://github.com/soma-project/soma-core.git
cd soma-core
pip install -e ".[dev]"

pytest -v --cov=soma --cov-report=term    # 760 tests, ~97% coverage

python -m soma                              # quickstart verification

python dash/server.py                       # API server (http://localhost:8765)

Project Structure

soma-core/
├── soma/                  # Core library
│   ├── __init__.py        # SOMA facade (zero-config entry)
│   ├── __main__.py        # python -m soma quickstart
│   ├── agent.py           # SOMA_Agent: pipeline orchestrator + awareness
│   ├── engine.py          # WisdomEngine: problem decomposition
│   ├── hub.py             # ActivationHub: bidirectional activation
│   ├── evolve.py          # MetaEvolver: reflection + auto-evolution
│   ├── embedder.py        # SOMAEmbedder: fastembed + ONNX encoding
│   ├── vector_store.py    # NumpyVectorIndex: faiss ANN search
│   ├── config.py          # Pydantic configuration models
│   ├── base.py            # Data models (Focus, MemoryUnit, etc.)
│   ├── abc.py             # Abstract base classes
│   ├── langchain_tool.py  # LangChain BaseTool wrapper
│   ├── law_discovery.py   # Autonomous law discovery from clusters
│   ├── retry.py           # LLM retry with exponential backoff
│   ├── plugin.py          # Entry-points plugin auto-discovery
│   ├── quality.py         # QualityEvaluator: reasoning output scoring
│   ├── analogy.py         # AnalogyEngine: cross-domain structural matching
│   ├── competitors.py     # Live competitor benchmark adapters
│   ├── analytics.py       # Usage analytics & benchmark storage
│   ├── benchmarks.py      # 5D benchmark engine
│   ├── wisdom_laws.yaml   # Default thinking framework (bundled)
│   ├── hub/
│   │   ├── _core.py       # ActivationHub: bidirectional activation + frame detection
│   │   ├── _conflict.py   # ConflictDetector: contradiction detection
│   │   ├── _frame_detector.py  # FrameAnchoringDetector: cognitive bias nudge
│   │   ├── _retriever.py  # MemoryRetriever: multi-path recall
│   │   ├── _scorer.py     # RelevanceScorer: weighted scoring
│   │   └── _ranker.py     # MMRRanker: diversity re-ranking
│   ├── multi_agent/       # Multi-Agent Collaboration
│   │   ├── registry.py    # AgentRegistry: expert registration + matching
│   │   ├── router.py      # ExpertRouter: 3-tier routing
│   │   ├── consensus.py   # ConsensusProtocol: vote/LLM/dialectic synthesis
│   │   └── evolve.py      # DistributedEvolver: independent evolution + weight merge
│   └── memory/
│       ├── core.py        # MemoryCore: unified memory facade + 3-state isolation
│       ├── episodic.py    # EpisodicStore: L1 episodic memory
│       ├── semantic.py    # SemanticStore: knowledge triples + causal graph
│       ├── skill.py       # SkillStore: learned patterns
│       ├── scene.py       # SceneStore: L2 scene blocks
│       ├── profile.py     # ProfileStore: L3 user profile
│       ├── capture.py     # CapturePipeline: auto L2+L3 extraction
│       ├── causal.py      # CausalGraph: causal chain reasoning
│       ├── consolidation.py  # ConsolidationEngine: memory dedup
│       ├── forgetting.py     # ForgettingEngine: Ebbinghaus decay
│       ├── external.py       # External knowledge import
│       └── search_utils.py   # FTS5 shared search utilities
├── dash/                  # Dashboard & API server
│   ├── server.py          # FastAPI (REST + SSE streaming + auth)
│   ├── providers.py       # LLM provider manager
│   └── frontend/          # Vue 3 dashboard UI (i18n: EN/ZH)
├── docs/                  # Documentation (EN + ZH bilingual)
├── tests/                 # 760 tests, ~97% coverage
├── examples/              # Usage examples
└── pyproject.toml         # Build config

Citation

@software{soma2026,
  title        = {SOMA: Somatic Wisdom Architecture},
  author       = {SOMA Project Team},
  year         = {2026},
  url          = {https://github.com/sunyan999999/soma},
  note         = {Apache 2.0},
  version      = {1.1.1},
}

License

Apache License 2.0. See LICENSE.

_{🧠 Five minutes to integrate. A cognitive kernel that remembers, reasons, collaborates, and evolves.}