Graphiti boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses Graphiti as a shipped graph-memory system, not only its Python API. Its intended use is to establish how information enters retained graph state, how model-dependent processing updates that state, and how later retrieval supplies application or agent context. The target includes the core library, database adapters, REST and MCP services, shipped agent-memory instructions, example consumers and evaluation code. It does not treat Graphiti as a general enclosing agent runtime. SRC-1: graphiti_core/graphiti.py, server/graph_service/routers/ingest.py, server/graph_service/routers/retrieve.py, mcp_server/src/graphiti_mcp_server.py, examples/langgraph-agent/agent.ipynb.

The target class is memory/knowledge/context-engineering system. Entity extraction selects prompts by episode source and requests a structured response from the configured language-model path. Ingestion retrieves previous episodes, resolves entities and facts, and saves episode data. This establishes model-dependent memory processing rather than a database wrapper alone. SRC-1: graphiti_core/utils/maintenance/node_operations.py, graphiti_core/graphiti.py.

return await generate_prompt_response( llm_client, prompt_name, legacy_prompt, context, clients=clients, response_model=ExtractedEntities, group_id=episode.group_id, ) --- graphiti_core/utils/maintenance/node_operations.py

The whole-system boundary includes material shipped paths for production, maintenance, admission and consumption of retained state. Initial tracing followed the core ingestion path through persistence and retrieval; the REST routes through their request-scoped client; the MCP entry point through its tool declarations and queue worker; and the LangGraph notebook through retrieval, prompt assembly and asynchronous ingestion of conversation content. The notebook is part of the target's shipped usage, not an excluded application merely because it calls the library. Its LangGraph scheduler and model SDK are external dependencies; the notebook's glue and memory-use decisions remain included. SRC-1: graphiti_core/graphiti.py, graphiti_core/driver/neo4j/operations/episode_node_ops.py, server/graph_service/main.py, server/graph_service/zep_graphiti.py, mcp_server/main.py, mcp_server/src/services/queue_service.py, examples/langgraph-agent/agent.ipynb.

The boundary also includes maintenance responsibilities: community rebuilds, saga summaries, triplet insertion, episode removal, graph clearing and group-scoped service operations. These are shipped responsibilities whether called by a maintainer or an assistant. Shipped instructions explicitly tell a consuming assistant to search before work, retain new information and apply retrieved preferences and procedures. These instructions establish declared use, not proof that an external assistant follows them. SRC-1: graphiti_core/graphiti.py, graphiti_core/utils/maintenance/graph_data_operations.py, server/graph_service/routers/ingest.py, mcp_server/src/graphiti_mcp_server.py, mcp_server/docs/cursor_rules.md.

• Always search first: Use the search_nodes tool to look for relevant preferences and procedures before beginning work. --- mcp_server/docs/cursor_rules.md

• Capture requirements and preferences immediately: When a user expresses a requirement or preference, use add_memory to store it right away. --- mcp_server/docs/cursor_rules.md

Evaluation is included as executable design evidence. The shipped evaluation CLI invokes baseline construction and candidate evaluation; its implementation ingests multi-session messages, serializes graph-building results, reloads baseline results and requests model judgments. This inclusion permits analysis of the measurement procedure, not a claim that the evaluator was run or that its scores establish effectiveness. SRC-1: tests/evals/eval_cli.py, tests/evals/eval_e2e_graph_building.py.

Frozen evidence and limits

The sole source is the acquired repository at full commit 964bb11d1e7020e3cc67fab244268b0736b6c219. The checkout HEAD was verified against that pin. The applicability cutoff is 2026-10-12, the supplied run date; it does not imply evidence about changes after the commit. Overall evidence is code-grounded, with README, specifications and tool instructions supplying doctrine/design. No target, test, example, model call or database service was executed.

External graph engines, language-model and embedding providers, reranking dependencies, MCP client hosts and LangGraph implementation are not frozen here. Their interfaces and Graphiti-side adapters are included; their internals and deployed behavior are not. This prevents conclusions about provider reliability, database-engine guarantees or host enforcement beyond what Graphiti's own code requests. Dependencies and supported integrations are declared in SRC-1: pyproject.toml, README.md, docker-compose.yml.

Proprietary Zep services, its Context Graph Engine, live hosted documentation, the linked paper, release binaries and actual production deployments are outside this source set. This prevents transferring claims about Zep's managed governance, latency or scale to this Graphiti revision. The source itself distinguishes the products. SRC-1: README.md.

Graphiti is the open-source framework for building temporal knowledge graphs at the core of Zep's context infrastructure. Zep manages context graphs at scale, providing governed, low-latency context retrieval and assembly for production deployments. --- README.md

Available but not yet closely inspected source files are not access gaps. The register covers the entire repository at the commit; initial reads below are not an allowlist. Backend-specific details, all prompt variants, test assertions and example compatibility require further inspection before claims about them. Historical notebook outputs were encountered during source discovery but are not used as evidence of operation of this pinned revision; their originating code/model/environment pins have not been established. No controlled experiment or pinned deployment trace is established by this boundary.

Top-level coverage

Every top-level tracked file or directory is classified below. Included directories name material paths and responsibilities; inclusion does not mean every file has already been read. The table distinguishes initial tracing from available material awaiting deeper inspection. Exclusions are deliberate functional exclusions, not unavailable repository evidence. All path entries refer to SRC-1.

Source register