OmniMemora

OmniMemora is a local LLM gateway with a desktop-GUI-controlled agent integration model. The current user control/display surface is the packaged OmniMemora Desktop app. http://127.0.0.1:18011 is the only product data entry when routing is enabled. :5173 is a legacy browser dashboard/dev surface and is not required by the desktop GUI.

Current Shape

Surface	Role	Status
OmniMemora Desktop app	GUI	Current user control/display entry
:5173	Legacy dashboard	Dev/legacy only; not required by current desktop GUI
:18011	Gateway	Only product data entry when routing is enabled
:8765	Go runtime	Internal memory plane

Agent -> Gateway (:18011) -> compile/recall/inject -> agent-configured LLM path
                               |
                               -> Runtime (:8765)

Cloud Split (Current Product)

Layer	Responsibility	Not Responsible
Cloudflare (doloclaw.com)	External domain entry, control-plane API/auth/tenant/billing/policy-access, candidate fetch entry	Cloud memory plane, cloud compile engine, /memory/* primary write/read/delete
Railway	Recommendation candidate snapshot/state storage, lightweight async aggregation jobs	/memory/* primary path, main compile path
Local (18011 + 8765)	Active/fallback execution truth, promotion-controlled active policy	Remote override of local active

Phase 8 external product name is DoloToken. Internal engineering name remains Token Intelligence Lite.

DoloToken starts from localhost/local proxy for cost and capacity control. This does not remove doloclaw.com, Cloudflare, or Railway from the product strategy; cloud DoloToken hosting remains a later deployment option after storage, traffic, privacy, and cost capacity are measured.

DoloToken can ship as a proprietary local CLI/local-proxy download. It must support online update checks and operator-pushed update notices through product-owned release metadata. During the unsigned beta stage, macOS users may need to allow the app or CLI manually in Privacy & Security / Gatekeeper; the product must state this clearly instead of implying a signed silent updater.

Frozen Truth

• OmniMemora Desktop app is the current user control/display entry.
• :5173 is legacy/dev-only and must not be treated as a required desktop GUI dependency.
• :18011 remains the only product data entry when routing is enabled.
• Runtime is internal only.
• OmniMemora is an agent path and optimization layer. It must not be described as owning or selecting an upstream model.
• Current OpenClaw and Claude Code usage is Anthropic-compatible and routes to the user's configured MiniMax M2.7 path when enabled.
• gemma4:26b is a local Ollama model option on this machine, not the current OmniMemora upstream model and not evidence of the real agent route.
• A /v1/models response is only a compatibility surface response; it is not proof of the user's actual agent model, upstream health, or product routing truth.
• By default, /v1/models lists no local Ollama model. OpenAI-compatible models appear there only when explicitly configured, and local Ollama entries must not be labeled as OmniMemora-owned.
• /metrics/core_capabilities is the current MVP savings truth surface (real_input_v1); /metrics/summary may return degraded no-historical-scan status.
• Adapter-to-runtime contract changes must pass contract tests.
• Agent control lives in the desktop GUI, not in agent self-selection.
• Agent integration is two-layer:
  • 使用 OmniMemora: high-frequency routing switch
  • 接入 OmniMemora: low-frequency install/uninstall switch
• Claude Code must not have OmniMemora MCP force-mounted as a global default. MCP may remain available as an explicit opt-in auxiliary surface, but it is not the product ingress, does not prove routing, and must not replace the desktop-controlled 接入 / 使用 flow.
• Codex 接入 OmniMemora prepares an OmniMemora-managed profile/launcher and must not rewrite the official ~/.codex/config.toml; legacy provider rewrites remain removable/restorable only for backward compatibility.
• When routing is off, requests may still enter :18011, but the gateway must stay in transparent passthrough mode.
• Agent detection must not auto-attach or auto-enable routing.
• Parent cards are the control granularity; temporary subagents are runtime-visible but not independent control cards.
• Pure local mode keeps cloud updates and usage reporting off by default.
• Enabling cloud policy updates implies minimal telemetry upload for policy quality improvement.

Current Phase

正式 roadmap phase：Phase 7（当前主线：Structured Compile MVP）（见 0_blueprint/ROADMAP.md）

Phase 标签说明：7_docs/internal/phase6/ 为 internal historical workstream，已收口并保留为治理/发布历史索引。当前产品能力工程入口是 7_docs/internal/structured_compile/README.md。 下一正式阶段：Phase 8（DoloToken / Token Intelligence Lite）已作为下一阶段固定在 roadmap，用于解释 token 花费、诊断浪费、推荐优化并证明实际节省。Phase 8 不得退化为普通 usage dashboard；当前阶段只做 workflow/token ROI 和 User Pattern Lite，不做用户画像。未来若升级为用户画像能力，必须作为新的显式产品阶段重新设计用户控制、隐私、存储和退出机制。

Start here:

• 0_blueprint/ROADMAP.md — 正式 roadmap SSOT
• 7_docs/internal/structured_compile/README.md — 当前结构化编译主线
• 7_docs/internal/token_intelligence/README.md — 下一阶段 DoloToken / Token Intelligence Lite 主线
• 0_blueprint/PRODUCT_DEFINITION.md
• 0_blueprint/PRODUCT_CONFIGURATION_AND_BOUNDARY_BASELINE.md
• 9_adr/ADR-0003-interface-access-paths.md
• 9_adr/ADR-0002-cloud-refactor.md

Quick Start

./start.sh

Healthy startup means:

• gateway health passes on :18011
• startup does not print false [OK]
• timeout or child failure exits non-zero
• detected agents may appear in the UI, but default control state stays off until the user enables it

Metrics Truth Surface

Endpoint	Role
/metrics/core_capabilities	Current MVP savings truth (real_input_v1)
/metrics/summary	Legacy/aggregate summary; may be degraded when historical scans are intentionally skipped
/proxy/status	Diagnostics
/compile/status	Diagnostics
/agents/live	Diagnostics
/agents/metrics	Diagnostics

If current MVP savings UI and historical summary disagree, trust /metrics/core_capabilities for real-input token savings.

Repository Layout

Path	Role
5_connectors/adapter/	Active gateway and adapter code
4_core/local-runtime/	Internal memory plane
6_console/demo-dashboard/	Dashboard
7_docs/internal/structured_compile/	Current structured compile mainline
7_docs/internal/token_intelligence/	Next Token Intelligence Lite mainline
7_docs/internal/phase6/plan/	Closed phase6 workstream and post-close governance records
5_connectors/archive/	Archived connector/plugin experiments

Non-Goals

• not a multi-entry product
• not a plugin-first product surface
• not a runtime-direct product
• not a second control plane beside the gateway
• not an auto-attach or silent-takeover product
• not a cloud-hosted primary memory plane

Governance

• 3_governance/AUDIT_SCHEME.md — 审计触发规则、执行骨架、结论路由
• 7_docs/internal/phase6/plan/OmniMemora_Cloud_Local_Sync_Check_2026-04-30.md — 云端-本地同步核对记录（2026-04-30）
• docs/spec/OMNIMEMORA_MVP_PROMO_2026-04-30.md — 闭源受控发布宣传文案草案（MVP）

Phase 标签说明：内部执行阶段标签（如 internal Phase 6 workstream）只表示执行 workstream。正式产品阶段以 0_blueprint/ROADMAP.md 为准。