Continuity synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From conversation to effects
Continuity is an enclosing desktop AI runtime, not merely a memory service. It constructs model requests, executes requested tools, supplies their results to subsequent calls and retains work for later use. Its declared ambition is a local-first workspace where conversation produces tasks, notes, decisions and persistent understanding. This account covers the shipped desktop, standalone MCP memory service, plugin host/SDK, organization server and both shipped plugins at commit 4ca8f6b4108aa4494e3861ed33c8019dbd662c67, with cutoff 2026-10-14. SRC-1 supplies inspected implementation, design instructions and attributed status claims; no target execution, supplied run trace or causal experiment supports this analysis. External providers, installed clients, actual vault contents and deployment behavior are outside the frozen evidence. Whole-system functional scope does not mean an exhaustive audit of every included path. See the boundary and source register and EPI-CLM-conversation-understanding.
A user send starts the foreground loop. Code loads the selected thread's saved role/content, builds a fresh system message from shipped guidance and available retained context, and calls a configured OpenAI or Anthropic model. The model proposes text and function calls; desktop code owns sequential dispatch and subsequent calls. Memory effects can be global while canvas work is thread-scoped. Missing credentials produce a configuration response. Context-loading failures can silently omit individual layers. The ten-iteration counter bounds only this foreground continuation loop: it does not terminate a hung tool, bound all system model calls or guarantee a saved final answer. A tool-producing last round can perform effects without a final answer round; exceptions do not roll back completed writes. These are source-defined mechanisms, not observed reliability claims. See RT-RTE-desktop-chat, RT-OBJ-current-messages and RT-CMP-ai-client.
Retention and later context
The main continuity mechanism is retained text in SQLite, with symbolic access and scheduling fields. Saved conversation is distinct from keyed facts, fact revision snapshots, caller-recorded learning observations and synthesized narratives. Human profile/custom-instruction saves are another admission mechanism. Desktop remember admits model-proposed global facts without per-item confirmation; MCP callbacks admit caller-supplied facts, observations, relations and narratives under schemas and SQL rules. Generic MCP caller identity does not establish human approval. These checks make content eligible for storage and retrieval, not warranted as true. See MEM-OBJ-messages, MEM-OBJ-memories, MEM-OBJ-memory-metadata, MEM-OBJ-memory-versions, MEM-RTE-desktop-retention, MEM-RTE-mcp-memory and MEM-RTE-profile-context.
Later sends receive facts by automatic supply and can request additional recall. Automatic selection takes the 50 most recently updated active memories across scopes, groups them by type and applies a nominal 4,000-character counter. It is not relevance selection against the current question or active project, and it is not a total-context cap. Source and project attribution are omitted from this passive display. Requested key recall, substring search and project retrieval can recover other retained content. A connected vault adds a separately delivered stored overview and requested file reads; its dedicated overview can bypass the general fact budget. These paths give text contextual or instructional influence through model interpretation, not deterministic compliance. See MEM-RTE-passive-context, MEM-RTE-desktop-retention, MEM-RTE-vault-context and RT-BAP-system-guidance.
Narrative maintenance adds a durable derived briefing. Scheduled checks require active memories and the documented narrative/learning conditions; a snapshot check can skip unchanged input, while manual initiation bypasses scheduling conditions. Synthesis receives all active memories across scopes, scoped pending observations and prior briefing/learnings—not raw saved conversation or the observations' source messages. Code automatically saves model output, including a raw-text fallback after JSON parse failure, then separately marks observations absorbed. Manual initiation is not approval of the generated content. The foreground later receives briefing text and displayed confidence/version/date, not its full structured reasons. Current-row replacement is wired, but faithful compression, entailed derivation and novel claim production cannot be distinguished without concrete premises/output evidence. See MEM-RTE-narrative-synthesis, MEM-OBJ-learnings, MEM-OBJ-narratives, EPI-RTE-narrative-generation, EPI-RTE-narrative-format and EPI-RTE-narrative-retention.
This maintenance has two consequential trust boundaries. First, synthesis instructions treat source=user as ground truth, although callers can supply that label and a desktop content update can retain an older label. Confidence is generated or caller-reported, not calibrated. Second, archiving a fact excludes direct active reads but does not retract its paraphrases from previous narratives, saved history or remote copies. Fact version history is not narrative rollback history. Typed contradiction/supersession links expose assertions but do not adjudicate them or automatically withdraw endpoints. Thus operational retention and use must not be read as evidence-consuming factual acceptance or lifecycle integration after acceptance. See MEM-CLM-source-trust, EPI-RTE-memory-withdrawal, EPI-RTE-relation-retention and EPI-RTE-narrative-consumption.
Delegated research and delivery
Deep research is the representative connected multi-step workflow. A conversation request or research-button hint still depends on foreground model selection of deep_research. After configuration checks, desktop code starts an independent pipeline and immediately returns initiation text. A planner proposes subquestions; code runs isolated investigators concurrently, selects sources and truncates each source to 2,000 characters. Only selected findings cross into later stages, not the full investigator contexts. These windows resemble task-local frames with returned values, but the returned text is copied content, not a live reference to a reloadable context. See RT-RTE-deep-research and its workflow account.
A reviewer receives 200-character finding prefixes and question statuses to propose further investigation. Code controls rounds and follow-up limits; the fifteen-read setting is a continuation threshold checked after current parallel work, not a strict total-read reservation. Failed review can proceed to synthesis. External Perplexity answers can enter as high-confidence findings without local factual verification; failed analysis can contribute 500-character source prefixes. Final synthesis receives 300-character findings plus a title/URL list, without each finding's original source association or confidence. Adequacy judgment controls further work, not truth acceptance or citation fidelity. See EPI-RTE-research-gap, EPI-RTE-research-continuation, EPI-RTE-provider-acquisition, EPI-RTE-excerpt-fallback and EPI-RTE-research-findings.
Delivery has separate milestones: engine result handoff, canvas append, summary display and summary persistence. The engine does not await the asynchronous desktop completion callback. Canvas failure can be followed by a summary claiming the report was written; the writes are not one atomic delivered unit. Cooperative cancellation cannot abort an in-flight request and can still deliver partial output through the same callback. These recovery paths preserve work rather than enforce successful evidence review. Later history supplies the saved summary, and the independently inspected read_canvas tool can return stored report blocks to a later foreground call. That positive consumer witness does not complete the workspace inventory or prove canvas input to personal narrative synthesis. See the workflow annotation, RT-RTE-deep-research, MEM-OBJ-workspace-state and report verification.
Other effect boundaries
Plugins and organization memory do not inherit a universal foreground policy. Plugin host handlers check capability strings, but authenticated registration copies the caller's own list. Shared credentials and broad native process/filesystem grants prevent interpreting those checks as independent least privilege or an OS sandbox. Host prompt/tool registration also does not establish delivery into the frontend's separate arrays. Organization server admission validates schema, authentication and timestamps, not truth. The shipped pull wrapper omits the database selector and therefore reaches the host's default app database rather than establishing admission to memory.db. Missing withdrawal markers and incomplete cursor/event wiring further prevent a joined, convergent desktop/server memory claim. See RT-RTE-plugin-admission, RT-CLM-plugin-controls, RT-RTE-org-ingestion, MEM-OBJ-org-memories and MEM-CLM-sync-withdrawal.
Threads Engagement separately schedules search, model persona-fit scoring and draft generation into plugin SQLite. A score threshold admits queue work; it does not validate post or reply truth. The review UI affords editing and selection, while external posting occurs through the approval endpoint, not the scheduler. That endpoint does not authenticate human review or enforce previous approved/rejected status as a publication veto; it can select the first available draft. Retained rejection/outcome records do not establish an update to future drafting guidance. See RT-RTE-threads-engagement, MEM-OBJ-threads-results, EPI-RTE-threads-selection and EPI-RTE-threads-publication.
Assessment at this boundary
The strongest learning-related contribution is an implemented trace-fed continuity path: conversational turns and tool feedback can feed automatic fact writes, facts reach later context/recall, and active trace-derived facts can also feed automatically retained narrative briefings. These are wired durable behavior-shaping updates; they establish neither faithful extraction nor improved capacity for future action. The absence of an inspected desktop writer for the separate learnings observation table does not negate these paths. Generic observation and prior-briefing origins remain unresolved. The documented developer memory bank prescribes task-summary updates and later reads, but remains doctrine rather than demonstrated desktop execution. See RT-RTE-desktop-chat, MEM-RTE-desktop-retention, MEM-RTE-narrative-synthesis, MEM-ABS-learning-extractor, MEM-RTE-developer-memory and profile verification.
Other properties remain separate. The accepted runtime does not establish formulated-theory routes sufficient to assess theory-builder conditions 1–4 individually: localized theory content, decisions dependent on that content, content-directed criticism and retained criticism shaping another round are each unestablished here, not evidenced absent from opaque model processing. Research follow-ups revise current investigation, not a demonstrated durable research method. No evidence shows criticism of a consumed theory improving future capacity. Reflection is not established by a briefing about the user: the required self-representation and both causal directions are not demonstrated. Automatic internal orchestration is wired, but an autonomous theory builder is not established by it. See RT-RTE-deep-research, RT-RTE-memory-admission and the epistemic bounded conclusion.
Retained facts and briefings can shape later operation, so they cannot be dismissed as merely external products. However, exercised self-improvement is not established: the records supply no independently specified improvement objective and no subsequent operation shown to causally depend on an evidence-responsive organizational update. Nor is there evidence of successful improvement. This is not an absence claim about all provider or workspace processes. See MEM-RTE-narrative-synthesis, RT-BAP-system-guidance and profile verification.
For conversational continuity, the source therefore supports multiple retained-context and correction mechanisms, with coarse automatic selection and incomplete derivative withdrawal. For open research, it supports delegated investigation and partial-work recovery, not certified findings or a durable-delivery receipt. For shared memory and publication, the actual database and endpoint boundaries constrain what integration and human control can be claimed. Concrete changes that would alter this assessment include authenticated provenance, evidence-consuming factual admission, derivative withdrawal, explicit memory-database routing and cursor/tombstone handling, and awaited delivery acknowledgments. Registered input/output traces could resolve transformation semantics and exercised context use; objective-bearing intervention comparisons would be needed for capacity-gain or causal-improvement attribution. These are evidence thresholds and system changes relevant to the findings, not adoption recommendations.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed execution or causal comparison | SRC-1; RT-RTE-desktop-chat, RT-RTE-deep-research, MEM-RTE-narrative-synthesis | Frozen source and shipped instructions; no target execution or registered run/experiment | Live activation, reliable delivery, usefulness, improved future capacity or causal self-improvement | Registered route-linked execution traces; independently specified objectives and appropriate intervention/outcome comparisons |
| Packaged distribution parity remains incomplete; carried report-verification limit | SRC-1; RT-RTE-plugin-admission, MEM-RTE-mcp-memory; verification limits | Source launch paths select committed host/SDK/server distributions; only targeted bundle inspection | Equivalence to a specific installed package's admission, discovery or memory delivery | Complete source/build provenance and comparison of selected distribution artifacts |
| Workspace inventory remains coarse and incomplete; carried report-verification limit | SRC-1; MEM-OBJ-workspace-state, RT-RTE-deep-research; verification limits | Research callback, journal/briefing paths and positive read_canvas witness inspected; full canvas/database and journal auto-link chains not traced | Exhaustive workspace read-back, context-budget and trace-fed-update comparisons; canvas-to-personal-synthesis attribution | Part-specific admission, maintenance and later-consumer records, including complete journal auto-link callers |
| Vault fidelity, containment and cleanup audit incomplete; carried report-verification limit | SRC-1; MEM-RTE-vault-context, EPI-RTE-vault-projection; verification limits | Formatter, index writes and delivery inspected; scanner/file containment and disconnect cleanup not fully audited; actual files outside evidence | Secure-path, complete-index and complete-withdrawal guarantees | Complete scanner/file-operation and disconnect traces/records, with frozen file inputs and fidelity checks |
| Relational integrity and organization event/search/auth integration incomplete; carried report-verification limit | SRC-1; MEM-OBJ-memory-links, MEM-OBJ-org-memories, RT-RTE-org-ingestion; verification limits | Endpoint checks, SDK routing and synchronization defects inspected; full integration not established | Complete relational-policy and end-to-end organization route comparisons; successful sync or system-wide absence of event delivery | Full mutation/integrity and event/search/auth producer-to-consumer records; registered end-to-end evidence |
| Source labels and generic input provenance/control remain unresolved | SRC-1; MEM-CLM-source-trust, MEM-OBJ-learnings, MEM-RTE-mcp-memory | Caller/default/inherited labels and observation-only synthesis inputs inspected | Authenticated human authorship/approval, source truth and complete original trace-source enumeration | Authenticated admission provenance, grounded source references and caller-control evidence |
| Model transformation semantics and confidence unvalidated | SRC-1; EPI-RTE-narrative-generation, EPI-RTE-research-findings, EPI-RTE-threads-drafting | Prompts, parsers and consumers inspected; no concrete premises/output trace or factual evaluation | Semantic preservation, entailed or ampliative production, calibrated confidence and factual acceptance | Route-linked inputs/outputs with claim-specific evidence, interpretation checks and intended-use acceptance criteria |
| Withdrawal and recovery do not cover every retained consequence | SRC-1; EPI-RTE-memory-withdrawal, MEM-OBJ-narratives, MEM-CLM-sync-withdrawal | Active-row archival, prior-narrative reuse, in-place narrative updates and ordinary sync payload inspected | Complete forgetting, narrative rollback and propagated remote withdrawal | Dependency-aware derivative withdrawal, retained narrative revisions and checked tombstone propagation |
| Provider state and external implementations are outside inspection | SRC-1; RT-CMP-ai-client, RT-CMP-ai-evaluator, RT-RTE-plugin-admission | Local request/launch adapters; remote weights, endpoint resolution, deployed grants, third-party services/clients unobserved | Immutable model identity, parameter fixity, complete provider-memory classification, external-client behavior and deployed isolation | Frozen provider/deployment contracts and corresponding implementation or registered operation evidence |
| Developer memory and arbitrary metadata alternatives unresolved | SRC-1; MEM-RTE-developer-memory, MEM-OBJ-memory-metadata | Developer doctrine and metadata container inspected; historical document derivation, update agency and arbitrary payload consumers not established | Exercised development-memory updates and complete form, authority or trace-fed-update inventory | Grounded document revision/loading evidence and payload-specific producer/consumer records |
The accepted reconciliation identifies no unresolved report conflict. Report verification declares the four carried local limits above; profile verification declares no additional limit. All comparison axes remain partial rather than asserting complete coverage. No source reads or new records were added for this synthesis.