Engraph boundary

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

Boundary and evidence

Engraph is in scope as a memory/knowledge/context-engineering system: a local Markdown-vault knowledge service that produces retained notes and derived indexes, maintains them, and returns searches and context to agents. It is not the enclosing conversational agent runtime. Its own retrieval issues embedding calls and optionally query-orchestration and reranking calls. The whole-system boundary includes the shipped binary and library, CLI callers, MCP and HTTP adapters, onboarding and identity, file synchronization, note-writing and migration paths, agent-facing skills, setup prompts and test definitions. These are parts of the target, not excluded applications merely because they call library functions (SRC-1, src/main.rs, src/lib.rs, src/serve.rs, src/http.rs, src/search.rs, README.md).

The intended use is to explain how a user's vault becomes persistent, searchable agent context, including the return path from agent-requested writing to later retrieval. The repository declares that purpose:

engraph turns your markdown vault into a searchable knowledge graph that any AI agent can query --- README.md

The implementation establishes model-dependent consumption, rather than only a static tool catalog:

let query_vec = embedder .embed_one(expanded_query) .context("embedding query")?; --- src/search.rs

Persistence, maintenance and reuse belong to the same shipped system. The CLI opens the stored vault and calls the writer; note creation writes Markdown and inserts chunks, vectors, full-text entries and edges. The server opens engraph.db, obtains its saved vault path, loads models and starts the watcher. Startup reconciliation runs the shared indexer before consuming filesystem events. Search handlers pass the shared store and loaded models to retrieval; context handlers return retained note content or assembled bundles (SRC-1, src/main.rs, src/writer.rs, src/indexer.rs, src/serve.rs, src/watcher.rs, src/context.rs). Identity extraction and formatting are also included: freshly indexed vault facts can subsequently supply the identity response (SRC-1, src/indexer.rs, src/identity.rs, src/http.rs).

let store = Store::open(&db_path)?; let config = Config::load()?; let embedder = crate::llm::LlamaEmbed::new(&models_dir, &config)?; --- src/serve.rs

let search_output = crate::search::search_internal(topic, 5, params.store, embedder)?; context_topic_from_results(params, topic, &search_output.results, max_chars) --- src/context.rs

The boundary is whole-system, not only the retrieval library or MCP interface. The table below classifies every top-level tracked file or directory at the frozen commit. “Included” makes the paths available to later source inspection; it does not claim that every implementation detail has already been inspected. Initial inspection traced entry points and the material state loop. Supporting module internals not yet read are identified explicitly, not treated as unavailable evidence.

All table classifications concern SRC-1 at the reviewed commit; paths in the table are commit-relative. The source register below supplies pinned navigation.

External dependencies include the vault and its user-supplied contents, local model files and their download hosts, llama.cpp bindings, SQLite/sqlite-vec, filesystem notification support, MCP clients and HTTP clients; optional Obsidian CLI delegation and ChatGPT tunnel/client setup are interface dependencies, not additional frozen implementations (SRC-1, Cargo.toml, README.md, src/serve.rs). Their vendored/runtime internals, actual user vaults, model weights, external agent conversation loops, and tunnel operations are outside this evidence set. This prevents conclusions about downstream answer quality, agent obedience, model accuracy, host scheduling or external service reliability. The repository's own integration code and suggested prompts remain inside the boundary.

The frozen revision is f9a95bc96accc792c02ee384d9e6bf768a88c8c8; applicability ends at 2026-10-14. HEAD was checked against the acquired revision and the checkout was inspected read-only. No second repository, live webpage, model artifact or execution trace was added. The overall tier is code-grounded: implementation can support source-level behavior, while documentation supports attributed intent. No target, tests or examples were executed. In particular, ignored integration tests and CI declarations do not establish a passing current build or retrieval effectiveness (SRC-1, tests/integration.rs, tests/write_pipeline.rs, .github/workflows/ci.yml). This is an execution-evidence limit, not a source-access gap.

Source register