REFERENCE / AGENT INFRASTRUCTURE
The landscape.
Memory, retrieval, execution, coordination.
Related problems. Different responsibilities.
01 / CATEGORY INDEX
Where Sophia fits.
A reading of public documentation captured in August 2026, not a live survey. Sophia does not replace most of these tools.
01Memory layers
Remembering across sessions: extraction, consolidation, and retrieval of durable memory about users, agents, and work.
- Mem0 ↗Memory layer with extraction and graph migration
- Supermemory ↗Context engine with profiles and connectors
- Letta ↗Stateful agents with memory blocks
02Knowledge-graph memory
Structure and time: entities, relationships, episodes, and temporal validity, so facts supersede each other instead of piling up.
- Cognee ↗Memory control plane over a knowledge graph
- Graphiti ↗Temporal context graph with episodes and lineage
- Zep ↗Managed memory platform built on Graphiti
03Retrieval and RAG frameworks
Getting the right context into the prompt: loaders, indexes, chunking, and query pipelines over your own corpus.
- LlamaIndex ↗Data framework for LLM applications
- Haystack ↗Open-source pipelines for retrieval and RAG
04Agent orchestration and runtimes
Control flow for multi-step work: graphs, crews, checkpoints, retries, and human-in-the-loop pauses.
- LangGraph ↗Agent graphs with persistence and time travel
- CrewAI ↗Role-based multi-agent crews
- Mastra ↗TypeScript agent framework with workflows
05Durable execution engines
Workflow state that survives crashes and retries: every step is journaled so a failed run resumes exactly where it left off.
- Temporal ↗Durable workflow engine, state captured at every step
- Inngest ↗Durable step functions, each step retries independently
- Restate ↗Journaled handlers plus per-key durable state
06Agent observability and evals
Seeing what agents actually did: traces of every call, scored evals, and production monitoring that make failures visible and debuggable.
- LangSmith ↗Agent tracing, monitoring, and evals
- Langfuse ↗Open-source LLM tracing and evaluation
- Braintrust ↗Evals with quality gates at release time
07Sandboxed agent execution
08Vector databases
Storage and similarity search at scale: embeddings indexed in seconds, queried in milliseconds, with hybrid retrieval and reranking.
- Pinecone ↗Managed vector database for production retrieval
- Qdrant ↗Open-source Rust vector search engine
- Chroma ↗Open-source search infrastructure for AI
09Tool access brokers and agent auth
Credentials agents can safely act with: managed OAuth per tool, user-scoped permissions, and tokens the agent itself never sees.
- Composio ↗Managed auth and tool calling across integrations
- Arcade ↗MCP runtime for user-scoped agent authorization
- Auth0 for AI Agents ↗Agent identity and token vault, some pieces still shipping
10Governed state layer
One authenticated local record binding memory, evidence, authority, coordination, and time: a control plane agents act through, documented mechanism by mechanism.
- Sophia ↗This project: the platform reference documents every mechanism
Owner-hosted by design. No managed cloud offering.
02 / DOCUMENTED CAPABILITIES
Inspect the comparison.
8 memory-adjacent projects. 21 responsibilities. “Not documented” is not a claim that a capability is impossible to build.
Open the full matrix
Scroll horizontally to compare all projects.
| Capability | Sophiagoverned state layer | Mem0memory layer | LangGraphagent runtime | Cogneememory control plane | Graphititemporal context graph | Zepmanaged memory platform | Supermemorycontext engine | Lettastateful agent platform |
|---|---|---|---|---|---|---|---|---|
| Memory coreWhere the category is already strong | ||||||||
| Cross-session persistenceState survives the context window and can be restored later. | Native | Native | Native | Native | Native | Native | Native | Native |
| Semantic or hybrid retrievalMeaning-based retrieval is a documented primitive rather than entirely application-supplied. | Native | Native | Native | Native | Native | Native | Native | Native |
| Entities and relationshipsKnowledge has durable graph structure beyond an embedding or message list. | Native | Partial | Not documented | Native | Native | Native | Native | Not documented |
| Temporal fact validityFacts retain validity intervals or explicit current-versus-superseded state, not merely timestamps. | Native | Not documented | Not documented | Partial | Native | Native | Native | Not documented |
| Source lineageAn extracted claim or memory can lead back to the raw episode or document that produced it. | Native | Partial | Not documented | Partial | Native | Native | Native | Not documented |
| Governed truthWhether remembered claims remain defensible | ||||||||
| Deterministic evidence validationClaim evidence is mechanically checked against the original source rather than accepted on model confidence. | Native | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Queryable contradiction stateDisagreement remains an inspectable state instead of being silently overwritten or merely ranked away. | Native | Not documented | Not documented | Not documented | Partial | Partial | Partial | Not documented |
| Human correction authorityA human correction structurally outranks later model inference and remains distinguishable from it. | Native | Not documented | Not documented | Not documented | Not documented | Not documented | Native | Not documented |
| Source freshness and driftThe system detects changed source content and can identify knowledge that may now be stale. | Native | Not documented | Not documented | Partial | Not documented | Not documented | Partial | Not documented |
| Governed actionWhether agents act through enforceable boundaries | ||||||||
| Authenticated agent identityThe runtime knows which agent is acting through an authenticated identity, not a caller-supplied label. | Native | Not documented | Not documented | Partial | Not documented | Not documented | Partial | Partial |
| Enforced data and tool scopeReads and actions are limited by runtime policy rather than prompt instructions alone. | Native | Partial | Partial | Native | Not documented | Not documented | Partial | Native |
| Human approval boundariesSensitive operations can pause at a declared gate and resume only after an explicit decision. | Native | Not documented | Native | Not documented | Not documented | Not documented | Partial | Native |
| Journaled, conflict-safe reversalMutation history is first-class and a prior state write can be safely reversed without erasing intervening work. | Native | Not documented | Partial | Not documented | Not documented | Not documented | Partial | Partial |
| Bounded delegationA parent can create a narrower child with explicit scope, lifetime, lineage, and revocation. | Native | Not documented | Partial | Not documented | Not documented | Not documented | Not documented | Partial |
| Agent operating layerState for continuity, coordination, and work | ||||||||
| Session recovery and handoffGoals, decisions, pending work, and context can be reconstructed after restart or agent change. | Native | Not documented | Partial | Not documented | Not documented | Partial | Partial | Partial |
| Correction acknowledgementEach agent session can be shown unacknowledged human corrections and record that it received them. | Native | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Durable coordination stateWork, waits, messages, ownership, and completion live in shared state outside any one prompt. | Native | Not documented | Partial | Not documented | Not documented | Not documented | Not documented | Partial |
| Structural code intelligenceCode symbols, imports, calls, and changes are queryable as durable structure rather than text chunks alone. | Native | Not documented | Not documented | Partial | Not documented | Not documented | Not documented | Not documented |
| Operator controlWhat stays on hardware the operator owns | ||||||||
| Self-hosted operationA documented route exists to run the core system on infrastructure you control. | Native | Native | Native | Native | Native | No, by design | Native | Native |
| Bundled local retrieval inferenceRetrieval inference runs locally without requiring a hosted model or separately operated model server. | Native | Partial | Partial | Partial | Partial | Not documented | Partial | Partial |
| Managed cloud offeringA vendor-operated hosted service. Sophia declines this deliberately; the daemon and its state stay on hardware the owner controls. Graphiti’s managed product is Zep, shown in its own column. | No, by design | Native | Native | Native | Partial | Native | Native | Native |
- ● Native
- Documented responsibility.
- ◐ Partial
- Narrower scope or application composition.
- — Not documented
- Not found in the reviewed documentation.
- ○ No, by design
- A deliberate product boundary.
Capability marks were first captured 22 July 2026 and last re-verified 11 August 2026; the Zep column was captured 19 August. This comparison preserves that dated research without claiming a new verification. A namespace is not authorization; CRUD history is not safe reversal; shared memory is not durable coordination; local-model compatibility is not bundled inference.
03 / SOURCES
Check the reading.
Official project documentation
- Sophia ↗
Self-reported: graded from this project’s own reference and verification notes, not from an independent review. Every other column is graded from that project’s public documentation.
- Mem0 ↗
Open-source overview, API, graph-memory migration, and self-hosting docs.
- LangGraph ↗
Persistence, stores, memory, time travel, and multi-agent documentation.
- Cognee ↗
Memory, knowledge graph, search, tenant isolation, and self-hosting documentation.
- Graphiti ↗
Temporal context graph, episodes, provenance, search, and deployment documentation.
- Zep ↗
Zep Cloud docs: context graph, facts and temporal validity, episodes, graph search, threads, and SDK reference.
- Supermemory ↗
Memory API, context, profiles, connectors, search, and local-operation documentation.
- Letta ↗
Memory blocks, archival memory, shared blocks, agent state, and self-hosting documentation.
Category descriptions were captured 19 August 2026. Product names belong to their respective owners. Corrections are welcome at chris@state-layer.ai.
Sophia's complete system reference ↗