GBrain boundary

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

Boundary and evidence

This run analyses GBrain as a memory/knowledge/context-engineering system serving existing agents. Its intended use is explicit memory with provenance, retrieval, correction and withdrawal, with optional synthesis and background enrichment. The target is the shipped repository artifact, not just its exported engine API or one MCP tool. The CLI, MCP transports, shipped agent instructions, context integration, maintenance and evaluation paths belong to the target. SRC-1 README.md, package.json, src/core/operations.ts.

Give the agent you already use a memory you control. GBrain stores explicit facts with their sources, supports corrections and withdrawal, and makes the same memory available across your agents. Start with keyless memory and keyword retrieval; add semantic search, synthesis, and background enrichment when you need them. --- README.md

The boundary kind is complete artifact, partial loop, not whole-system. Initial inspection establishes shipped entry points and representative admission, read, context-consumption, maintenance and evaluator paths. It does not exhaustively trace every state producer, persistence owner, background driver, host adapter or evaluator. Those available but uninspected paths are not access gaps or excluded applications. They remain part of the registered artifact and may supply later evidence at the same commit. This boundary prevents an initial claim of complete loop coverage; it does not restrict later inspection to the paths listed in the source register.

Entry points and initial responsibility trace

• CLI and library: package.json exposes src/cli.ts as the executable and src/core/index.ts as the library entry, with additional engine, operations, search, synthesis, ingestion, AI and optimizer exports. The executable dynamically loads src/cli/main.ts. The inspected operation registry assembles page, search, graph, memory, job, skill, extraction and other handlers. Same-repository command drivers are included, not treated as external library consumers. SRC-1 package.json, src/cli.ts, src/core/operations.ts.
• Capture and admission: the CLI capture branch chooses remote MCP, local-owner delegation or an in-process operation handler. The registered capture handler submits a page mutation. The inspected admission function prepares a request, calls admitWrite, and waits for a receipt. This establishes a shipped state-admission path, not proof that every publication or crash-recovery path succeeds. SRC-1 src/commands/capture.ts, src/core/ops/pages.ts, src/core/persistence/page-mutations.ts.

return submitPageMutation(ctx, { operation: 'capture', params: p }); --- src/core/ops/pages.ts

const response = throwIfHeld(pendingAwareResponse(ctx, await waitForWrite(ctx.engine, row, ctx.config, waitMs())), false); --- src/core/persistence/page-mutations.ts

• Retained-memory reads: recall is registered as a read operation with entity, question, query, source and token-budget parameters. Its inspected handler resolves source scope and local-versus-remote visibility before reading eligible facts; bounded searches also identify its engine fact-query call sites. The full packing and retrieval implementation has not yet been inspected. SRC-1 src/core/ops/facts.ts.
• Model-facing consumption: the shipped query skill instructs an agent to retrieve, read, synthesize and cite stored content. This is an instruction contract, not evidence of compliance. The OpenClaw plugin is explicitly shipped in package.json; its registration creates the repository's context engine. The inspected assembly branch combines live context, core-memory additions, checkpoint information and resolved entity pointers into systemPromptAddition. The host still owns the enclosing conversation runtime. SRC-1 skills/query/SKILL.md, package.json, src/openclaw-context-engine.ts, src/core/context-engine.ts.

1. Read top results. Read the top 3-5 pages from gbrain to get full context.
2. Synthesize answer with citations. Every claim traces back to a specific page slug. --- skills/query/SKILL.md

systemPromptAddition: parts.join('\n\n'), --- src/core/context-engine.ts

• Maintenance: runMaintain reads health and doctor reports, invokes stale extraction and source-cycle maintenance, then emits before/action/after results. Dream-cycle internals and scheduled drivers remain to be traced. The separate capture skill is included as shipped ingestion doctrine; its explanatory text does not override the current executable admission path. SRC-1 src/commands/maintain.ts, skills/capture/SKILL.md.
• Evaluation: the inspected BrainBench command loads a corpus, refuses an empty fixture set, and calls runBrainBench with harness, suite, holdout, model-use and budget settings. The harness internals and adapters are available but not yet inspected; this establishes an evaluator entry point, not validated quality scores. SRC-1 src/commands/eval-brainbench.ts.

Functional inclusions and inspection limits

The tracked-tree inventory was inspected at the frozen HEAD. The following is an artifact map, not a whole-system coverage claim. Grouped rows account for every top-level tracked entry; inspection status distinguishes initial source reading from inventory-only visibility.

External host runtimes, their user configurations and private brain corpora are excluded unless their behavior is represented in the frozen repository. GBrain's shipped integration code and instructions are included, but OpenClaw, Hermes, Claude Code, Codex and other hosts are not themselves analysed. This prevents attributing host scheduling, conversation persistence, compaction or model compliance to GBrain. SRC-1 README.md, src/openclaw-context-engine.ts, src/core/context-engine.ts.

ownsCompaction: false, // delegate to legacy runtime --- src/core/context-engine.ts

Bun, PGLite/Postgres, model-provider SDKs and services, and optional connector infrastructure are external dependencies; only their shipped declarations and integration code at this pin are evidence here. This prevents independent claims about provider internals, live service availability or deployment compatibility. SRC-1 package.json, README.md.

The immutable boundary is commit da434385572e0dc0d825649179aabed4107d6e60 of https://github.com/garrytan/gbrain; the local checkout's HEAD matched that commit and its working tree was clean. The applicability cutoff is the opening run date, 2026-10-11. Later commits and live linked repositories/pages are outside the evidence set. No target execution, service calls or experiments were performed. The overall tier is code-grounded, with doctrine separately labelled. There is no inspected observed-run or causal-experiment evidence. Deployment numbers and performance claims in the README are only reported operation, not verified measurements. SRC-1 README.md.

Source register