Mem0
Hosted cloud API - zero setup, zero infrastructure. If you need memory in 5 minutes and don't care about governance, privacy, or local-first, Mem0 is faster to integrate.
Evidence
Every number was measured on local consumer hardware - mostly a Dell Inspiron 3582 running Zorin OS. No cloud instances. No GPU acceleration. No cherry-picking. Last verified July 29, 2026.
| Metric | Value | Context | Hardware |
|---|---|---|---|
| Skill retrieval latency | <1ms | vs 200ms-2s for LangGraph/AutoGen LLM planning | Dell Inspiron 3582, Zorin OS |
| Cross-domain routing accuracy | 100% | Trading to Governance to Analysis, no LLM reasoning pass | Dell Inspiron 3582 |
| Concurrent skills loaded | 509 | 28ms bootstrapping overhead | Dell Inspiron 3582 |
| Memory retrieval (406K memories) | <5ms | SQLite + Rust FFI, full scan | Dell Inspiron 3582 |
python scripts/system_capability_test.py --suite routing
python scripts/system_capability_test.py --suite routing
python scripts/system_capability_test.py --suite hotload
python scripts/system_capability_test.py --suite memory
| Metric | Value | Context | Hardware |
|---|---|---|---|
| HNSW vector search | 0.26ms | 16,219 embeddings, disk-persisted index | Consumer laptop |
| FTS5 full-text search (1K memories) | 2.6ms median | Phrase-first ranking, SQLite FTS5 | Consumer laptop |
| Session record per turn | 0.5ms | SQLite store, ~3ms total overhead per conversation turn | Consumer laptop |
| Session recall (20 turns) | 0.6ms | Tag-filtered search + client-side sort | Consumer laptop |
python -m pytest tests/unit/test_hnsw_index.py -v
python -m pytest tests/unit/test_sqlite_backend.py -k fts -v
python -m pytest tests/unit/test_session_recorder.py -v
| Metric | Value | Context | Hardware |
|---|---|---|---|
| Rust AVX2 GEMV (256x256) | 563us | 12.5x speedup over Python scalar (5.6ms) | Consumer laptop, AVX2 |
| Ternary kernel (12-layer model) | 67ms/token | vs 873ms/token Python - real-time vs noticeable lag | Consumer laptop, AVX2 |
| PredictiveCoder (dim=128) | 19x speedup | Rust PyO3 vs Python scalar | Consumer laptop |
| Batch cosine similarity | AVX2 accelerated | N-parallel vector comparisons in single SIMD pass | Consumer laptop, AVX2 |
cargo bench -p wm-neuro --bench ternary
cargo bench -p wm-neuro --bench predictive
| Metric | Value | Context | Hardware |
|---|---|---|---|
| MCP tool dispatch (median) | 29-33ms | 3-10x faster than typical MCP implementations (29-33ms vs 100-300ms) | Consumer laptop |
| MCP tool dispatch (P95) | 36-55ms | 95th percentile latency | Consumer laptop |
| MCP tool dispatch (P99) | 38-86ms | 99th percentile latency | Consumer laptop |
| Success rate | 100% | Across all tool calls | Consumer laptop |
| Memory per call | 0-0.18MB | Heap allocation per dispatch | Consumer laptop |
| Throughput | 29.38 req/s | Sustained requests per second | Consumer laptop |
| Homeostatic loop overhead | 0.35ms | Full loop including physical checks | Consumer laptop |
| Rate limiter pre-check | 452K ops/s | Rust EventRing, zero allocation hot path | Consumer laptop |
python scripts/benchmark_gauntlet.py --suite mcp
| Metric | Value | Context | Hardware |
|---|---|---|---|
| HLL cardinality tracking | 508KB / 100K memories | 1.2% error rate, O(1) space - runs forever without degradation | N/A (mathematical) |
| Count-Min Sketch | Sub-1KB | Frequency estimation with bounded error | N/A (mathematical) |
| Thompson sampling bandit | Beta posterior updates | Contextual tool selection - agent gets better with use without retraining | N/A (algorithmic) |
| Metric | Value | Context | Hardware |
|---|---|---|---|
| Memories indexed | 406,836 | SQLite + Rust FFI, sub-5ms retrieval maintained | Dell Inspiron 3582 |
| Events processed (Live Era) | 33,297 | voice_expressed, memory_created, oracle_cast, pattern_detected | Dell Inspiron 3582, Zorin OS |
| Memories (v15.8 era) | 111,665 | 2.2 million associations | Dell Inspiron 3582 |
| Current memories (v24.0) | 89,500+ | 14 galaxies, HNSW + FTS5 dual search | Consumer laptop |
| Test suite | 8,268 passing | 56 skipped, 0 failing, ~120s runtime | Consumer laptop |
cd core && python -m pytest tests/ --ignore=tests/archive_v14 --ignore=tests/archive_v11 --ignore=tests/archive --ignore=tests/archive_polyglot --ignore=tests/legacy --ignore=tests/adhoc --ignore=tests/verify -q --timeout=30
WhiteMagic is not the right tool for every job. Here is an honest comparison against Mem0, Letta, and standard RAG - including where competitors have genuine advantages.
| Capability | WhiteMagic | Mem0 | Letta | Standard RAG |
|---|---|---|---|---|
| Memory architecture | 6D holographic coordinates (temporal, semantic, emotional, relational, importance) | Flat vector store with metadata tags | Agent state + memory blocks, harness-managed | Linear vector embeddings, no relationships |
| Persistence horizon | Global - memories persist forever, rotate through galactic lifecycle zones | Session or conversation-scoped | Persistent within a harness deployment | N/A - stateless retrieval |
| Search modalities | HNSW vector + FTS5 full-text + graph walking + HRR projection + 6D coordinate lookup | Vector similarity only | Vector + keyword within harness | Vector similarity only |
| Memory relationships | Cross-galaxy associations, knowledge graph, 2.2M associations in v15.8 era | No relationship tracking | Limited - within agent context window | No relationships |
| Decay & consolidation | 8-phase dream cycle: triage, consolidation, serendipity, governance, narrative, kaizen, oracle, decay | No decay or consolidation | No dream cycle | No decay |
| Emotional encoding | Emotional valence dimension in 6D coords, 20+ emotional gardens (joy, grief, courage, love) | No emotional encoding | No emotional encoding | No emotional encoding |
| Capability | WhiteMagic | Mem0 | Letta | Standard RAG |
|---|---|---|---|---|
| Ethical governance | Dharma rules engine - YAML-driven, 4 profiles, hot-reloadable, 8-stage dispatch pipeline | Prompt-based guardrails | Harness-level tool permissions | None |
| Audit trail | Karma ledger - SHA-256 Merkle-chained, XRPL-anchored, append-only | Basic logging | Harness logs | None |
| Input sanitization | Shell injection detection, content scanning, internal field stripping | No | No | No |
| Rate limiting | Rust EventRing - 452K ops/s pre-check, zero allocation hot path | No | Harness-level | No |
| Circuit breaker | Yes - automatic failure detection and recovery | No | No | No |
| Sandboxing | 5-tier sovereign sandbox: thread, namespace, container, microVM, WASM | No | No | No |
| Capability | WhiteMagic | Mem0 | Letta | Standard RAG |
|---|---|---|---|---|
| Skill retrieval | <1ms (Muscle Memory) | N/A | 200ms-2s (LLM planning) | 50-500ms (vector search) |
| Vector search | 0.26ms (HNSW, 16K embeddings) | 50-200ms (cloud API) | 10-50ms (local) | 50-500ms |
| Local-first | Yes - by default. No cloud required. | No - cloud API | Partial - harness-dependent | No - cloud-dependent |
| Offline operation | Full operation including inference (ternary kernels, Ollama integration) | No | Partial | No |
| Polyglot acceleration | 7 languages: Rust, Go, Zig, Haskell, Elixir, Julia, Koka | Python only | Python only | Python only |
| Probabilistic tracking | HLL + Count-Min Sketch - O(1) space, runs forever | No | No | No |
| Capability | WhiteMagic | Mem0 | Letta | Standard RAG |
|---|---|---|---|---|
| Consciousness primitives | Citta stream, coherence metrics (8 dimensions), Smarana practice, presence quality | No | Agent identity within harness | No |
| Self-calibrating forecasts | Brier-scored predictions, 38 validated prescience claims, 940 points | No | No | No |
| Tool selection learning | Thompson sampling bandit - Beta posteriors converge with use, no retraining | No | No | No |
| Metacognitive awareness | Cardinality-velocity surprise gate - modulates learning rate by novelty | No | No | No |
| Session continuity | Session recorder with progressive recall, selective replay, FTS5 search | No | Agent state persistence | No |
| Dream cycle | 8-phase: triage, consolidation, serendipity, governance, narrative, kaizen, oracle, decay | No | No | No |
We do not pretend WhiteMagic is perfect for every use case. Here are the areas where competitors have real, honest advantages.
Mem0
Hosted cloud API - zero setup, zero infrastructure. If you need memory in 5 minutes and don't care about governance, privacy, or local-first, Mem0 is faster to integrate.
Mem0
Established market presence and documentation. WhiteMagic's docs are thorough but technical; Mem0 has a polished developer experience for quick starts.
Letta
Deep harness integration. Letta manages the agent runtime itself - tool calling, context windows, model routing. WhiteMagic is a substrate that any agent plugs into.
Letta
Memory models (June 2026) - specialized models for creating durable context that transfers across LLMs. WhiteMagic doesn't yet have model-trained memory compression.
Standard RAG
Simplicity. If you just need 'search a PDF and ask questions,' RAG is simpler and well-understood. WhiteMagic's 860 tools and 8-stage pipeline are overkill for basic retrieval.
Standard RAG
Ecosystem. LangChain, LlamaIndex, and the broader RAG ecosystem have thousands of integrations. WhiteMagic has 860 tools but a smaller integration surface.
Choose WhiteMagic if you need local-first operation, ethical governance as architecture (not prompts), memory that persists and consolidates, consciousness primitives for continuous agents, or you are building for air-gapped / regulated environments.
Choose Mem0 if you need a hosted memory API in 5 minutes and do not need governance, privacy guarantees, or local-first operation.
Choose Letta if you want a complete agent harness that manages tool calling, context windows, and model routing - and you are okay being harness-locked.
Choose standard RAG if you just need to search documents and ask questions. It is simpler, well-understood, and has a massive ecosystem.