Continuity boundary

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

Boundary and evidence

This run examines Continuity as a shipped desktop AI workspace, not just its standalone memory API. Its intended use is conversation-driven work with persistent personal and project context. The repository also ships a standalone MCP memory server, a plugin host and SDK, two plugins, and an organization memory HTTP server. These are included even where they run in separate processes. The target class is enclosing runtime: the desktop chat provider constructs model inputs, issues model calls, dispatches tools, feeds results into subsequent calls, and saves conversation output. Memory/knowledge/context engineering is a major responsibility within that runtime, not the entire target. The boundary kind is whole-system. These classifications follow SRC-1, app/layout.tsx, providers/chat-provider.tsx, server/index.ts, plugin-host/src/index.ts, plugins/continuity-threads-engagement/src/index.ts and org-server/src/app.ts.

The frozen evidence is repository https://github.com/uziiuzair/continuity at 4ca8f6b4108aa4494e3861ed33c8019dbd662c67. The acquired checkout's HEAD matched that revision and its tracked working tree was clean on inspection. Applicability is limited to that revision with analysis cutoff 2026-10-14, the supplied run date. No other repository, published package, live service, installed user data or revision is evidence. The overall tier is code-grounded: executable sources and shipped callers were traced, with documentation used as declared intent rather than proof of operation.

Classification anchors

The product declares its intended use in SRC-1, README.md:

A local-first AI workspace where chat is the write path and structure emerges from conversation. --- README.md

The enclosing model loop is implemented in SRC-1, providers/chat-provider.tsx:

// Tool calling loop - keep calling until no more tool calls let currentMessages = [...aiMessages]; let finalContent = ""; let response; const MAX_TOOL_ITERATIONS = 10; // Safety limit --- providers/chat-provider.tsx

The standalone memory service is not the sole shipped consumer. The same chat provider constructs each request using retained context (SRC-1, providers/chat-provider.tsx):

const parts = [ PERSONALITY_SYSTEM_PROMPT, ...(userProfilePrompt ? [userProfilePrompt] : []), ...(narrativePrompt ? [narrativePrompt] : []), ...(memoryContextPrompt ? [memoryContextPrompt] : []), WORKSPACE_AGENT_PROMPT, --- providers/chat-provider.tsx

The shipped behavioral prompt directs model-mediated retention, rather than leaving all memory production to an unspecified host (SRC-1, lib/ai/workspace-agent-prompt.ts):

On each conversation: 1. Check your pre-loaded memories for relevant context 2. Use `recall` to search for specific memories if needed 3. Use `remember` to store important new information --- lib/ai/workspace-agent-prompt.ts

The application also owns a model-dependent maintenance path (SRC-1, lib/ai/narrative-synthesis.ts):

const messages: ChatMessage[] = [ { role: "system", content: SYNTHESIS_SYSTEM_PROMPT }, { role: "user", content: userMessage }, ];

const response = await aiClient.chat(messages); --- lib/ai/narrative-synthesis.ts

Functional inclusions and traced paths

• Conversation production and later reuse: app/layout.tsx mounts ChatProvider; providers/chat-provider.tsx reloads saved thread messages, builds a fresh system prompt, runs the bounded streaming/tool loop, dispatches built-in, MCP and plugin tools, and saves final messages through lib/db/messages.ts. The same provider schedules a synthesis check after saved conversation activity. These are shipped driver responsibilities, not excluded external application behavior. SRC-1, app/layout.tsx, providers/chat-provider.tsx, lib/db/messages.ts.
• Memory admission, persistence, retrieval and maintenance: in-app memory tools reach lib/db/memories.ts, which writes versioned memories and archives rather than physically deleting through its forget path. lib/db/memory-db.ts opens and bootstraps the memory database. server/index.ts registers memory, project, search, relationship, lifecycle and narrative/learning tool modules for external MCP callers. lib/ai/memory-tools.ts loads retained memories into the chat prompt. lib/ai/narrative-synthesis.ts gathers memories, unabsorbed learnings and a previous narrative, calls the configured AI client, saves a narrative and marks learnings absorbed. providers/memories-provider.tsx reads retained state for the UI and supplies launch/manual synthesis paths. Inclusion establishes ownership of these paths, not the correctness of every write or scheduling condition. SRC-1, those paths and server/tools/narrative-tools.ts.
• Other retained work and context: canvas/database tools and their persistence providers, journal entries and links, briefing projections, and Obsidian vault scanning/read/write are included. Vault scan callers reach lib/vault/memory.ts; its retained index is later consumed by lib/ai/obsidian-tools.ts and the chat prompt builder. Deep research has its own model prompts and orchestration; the chat caller handles completion, canvas output and saved chat summary. SRC-1, components/settings/panels/VaultPanel.tsx, lib/vault/memory.ts, lib/ai/obsidian-tools.ts, lib/ai/research-engine.ts, providers/chat-provider.tsx, lib/db/journal.ts, lib/db/journal-links.ts, lib/services/briefing-service.ts.
• Plugin and organization paths: the frontend manager spawns the plugin host and plugin processes; host handlers expose database access, prompt/tool registration, events, settings, MCP controls and UI registration. The SDK and both shipped plugins are included. Org Memory Sync subscribes to memory events, queues pushes, pulls server records through SDK database writes, and provides a search tool and prompt. The organization server supplies push, pull and search endpoints. Include both ends and the SDK default database routing when assessing whether they actually form the intended memory loop; do not assume integration from their names. SRC-1, lib/plugins/manager.ts, plugin-host/src/handlers/db.ts, plugin-sdk/src/db.ts, plugins/continuity-org-memory-sync/src/index.ts, plugins/continuity-org-memory-sync/src/sync-engine.ts, org-server/src/app.ts.
• Independent plugin model workflow and evaluators: Threads Engagement owns scheduled search, model relevance scoring, model reply drafting, a plugin-local database, a review UI and an approval endpoint that calls its Threads API wrapper. Its AIEvaluator is a production relevance/drafting component, not an experimental evaluation of Continuity memory quality. Organization-server tests include an injected-server push/pull/search round trip; their text is implementation evidence of a test definition, not an observed passing run or proof that the desktop sync integration works. SRC-1, plugins/continuity-threads-engagement/src/index.ts, plugins/continuity-threads-engagement/src/cron.ts, plugins/continuity-threads-engagement/src/ai-evaluator.ts, plugins/continuity-threads-engagement/src/ui-server.ts, org-server/tests/roundtrip.test.ts.
• Shipped instructions: runtime prompts in lib/ai/, plugin-injected prompts, and developer memory-bank maintenance instructions in CLAUDE.md/claude_docs/ remain within the evidence boundary. The developer instructions describe a separate agent-assisted development procedure, not automatic in-app execution. Source documentation and progress claims require implementation corroboration before becoming claims about this frozen runtime. SRC-1, CLAUDE.md, claude_docs/ACTIVE_CONTEXT.md, lib/ai/workspace-agent-prompt.ts, plugins/continuity-org-memory-sync/src/index.ts.

Repository coverage

Every top-level tracked file or directory is accounted for below. All paths in the table refer to SRC-1. “Included” fixes functional scope; it does not claim exhaustive line-by-line inspection. Inspection is concentrated on entry points, state producers, admission/persistence, maintenance and later consumers. Available ancillary files not yet inspected are identified separately from unavailable evidence or deliberate exclusions.

External dependencies and evidence limits

The implementation delegates model generation to configured OpenAI or Anthropic clients, search to configured web/search providers including Tavily and Perplexity, and host operations to Tauri plugins and SQLite implementations. Threads Engagement additionally relies on the Threads API. External MCP servers, third-party plugins, installed external MCP clients such as Claude Code/Cursor, actual Obsidian vault files and deployment infrastructure are outside this repository evidence. Their integration code is included; their implementations, data and operation are not. Consequently this run cannot establish provider quality, external-tool policy, external-client prompt/reload behavior, successful cross-client use, external vault content correctness or live deployment reliability. SRC-1, lib/ai/index.ts, lib/ai/openai.ts, lib/ai/anthropic.ts, lib/ai/perplexity.ts, lib/ai/web-tools.ts, lib/mcp/stdio-transport.ts, lib/mcp-auto-setup.ts, plugins/continuity-threads-engagement/src/index.ts, src-tauri/src/lib.rs, server/db/connection.ts.

There is no unavailable repository source identified at this boundary. Files available but not inspected remain distinguishable from access gaps in the coverage table. No target code, tests, examples, model providers or services were executed. No supplied execution trace or controlled experiment is registered. Test definitions and source-native status claims therefore cannot establish an observed passing run, end-to-end integration success, memory usefulness, calibrated confidence, comparative performance or causal improvement. The boundary is complete as a source-defined functional scope; it is not a completed audit of every included path.

Source register