Origin synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From supplied material to durable state
Origin, branded Wenlan in the reviewed checkout, is a local knowledge service that retains captures and imported material, organizes them into memories and pages, and supplies context to later human and agent requests. It is not the enclosing conversational agent. This account covers the shipped core, owning daemon, desktop/frontend, CLI, MCP transports, client instructions and optional relay at commit 61b462ea6adf019b3e27d3f58ac5a0fe055f76b1, with an applicability cutoff of 2026-10-15. Its evidence is SRC-1 implementation and attributed doctrine/design, not target execution. No program, test, model or service was run; no independent observed-run or causal evidence establishes activation or benefit. The boundary and Source register distinguish included responsibilities from deployment dependencies and deliberate exclusions. This synthesis uses the supplied members without additional target-source reads.
The daemon owns durable state and background work. Desktop, CLI and MCP are callers; external host agents interpret the shipped skills. An ordinary MCP capture sends content and metadata to the store handler. Code resolves space and trust, checks admission and returns a persisted receipt separately from enrichment status. Capture can work without a configured generative provider; embedding infrastructure remains a dependency. A receipt does not promise eventual organization. Reduced-trust superseding writes are staged, but appends do not share that predicate and unidentified local writes receive full trust. This is selective change control, not universal human approval or authenticated authorship. The optional quality gate checks capture admissibility, including noise and credentials, not source truth; configuration can disable it (RT-RTE-capture-store, EPI-RTE-quality-check).
Conversation and document imports have separate parsing, deduplication and synchronization paths. They are not established to share the server capture gate. Retention is also plural: libSQL holds prose, source text, structured fields and control metadata alongside vectors; optional redacted episode rows preserve an input alternative, while fact-child retrieval rehydrates the parent. These representations do not constitute independently accepted facts merely because they are stored (MEM-RTE-import-inputs, MEM-OBJ-memories, MEM-OBJ-episode-rows, MEM-OBJ-fact-children).
Organization, compilation and evidence support
Automatic enrichment is deferred to resource- and configuration-gated scheduler slices. Durable cursors and versioned attempt receipts guide continuation; selected models propose classification, extraction or titles, and code admits derived state. Recent-memory recaps and optional summary rollups create additional retained representations. A recap can retain raw burst context separately from its generated summary; a rollup root is generated from bucket titles rather than all original bodies. These distinctions matter to what later consumers can know. The scheduler's atomic one-call bound covers shared provider handles within an automatic slice, not all daemon inference, foreground rebuilds or provider internals (RT-RTE-ambient-work, MEM-RTE-enrichment-slices, MEM-RTE-burst-recaps, MEM-RTE-summary-rollup).
Page refresh selects linked sources, caps each snippet to 800 characters and combines them with task guidance. The model returns candidate prose, not mutation authority. Citation processing rejects zero verified occurrences or a majority unverified. Here verification means lexical overlap, with paragraph fallback; it does not establish entailment, source truth or coverage of uncited propositions. The retained implementation deduction shows that removing short tokens can admit a sentence that negates its source. Ordinary machine updates stage revisions for human-owned pages. Forced rebuild can override user-edited protection but still stages authored pages. Concurrency protection is branch-specific: forced refresh without candidate sources carries source revision and page incarnation into the body write; ordinary and candidate-source refresh carry the source counter without that incarnation. Neither the citation gate nor these fences apply universally to every writer, including external Markdown edits (RT-RTE-page-refresh, EPI-RTE-citation-check, RT-RTE-filesystem-edit).
Origin also implements a distinct, stronger evidence-consuming support pipeline, called M5. A separate daemon task extracts page sentences and asks an eligible, snapshot-restricted on-device judge whether selected linked evidence directly entails each claim. It validates item identities and finite scores; the inspected production threshold is 0.75. Publication separately requires current eligible support, accounted-for candidates and a completed/current run, then recomputes the outcome transactionally. This is an implemented acceptance criterion for current page-version evidence support, with retained support edges and subsequent operational use. It is not proof of source truth, model accuracy, exhaustive evidence search or production of an accepted novel proposition. Actual input/output witnesses are missing (EPI-RTE-claim-extraction, EPI-RTE-claim-judgment, EPI-RTE-support-publish).
Support has conditional force. At truth generation zero, exposure passes through. After cutover, judged unsupported pages without human review are hidden from automatic page consumers, but unevaluated pages remain visible. Human-reviewed pages override machine support; collection grants permit reduced entries and named-page grants permit bounded full inspection. A version-bound human-review mark is separate from installing a revision. Evidence changes can withdraw support and enqueue rechecking without asserting the opposite claim or hiding all prose. Thus the mechanism regulates reliance under a selected exposure regime; it is not a rule that only verified knowledge can be read (EPI-RTE-truth-exposure, EPI-RTE-truth-cutover, EPI-RTE-page-attestation, EPI-RTE-support-withdrawal).
Other review actions retain narrower meanings. Dismissing a memory revision keeps an independent capture; dismissing a manufactured page card deletes that proposal; accepting a contradiction refinement can resolve its queue item without changing either fact. Lint produces advisory findings, repair apply authorizes a manifest-bound mutation, and later verification checks operation-specific postconditions. None is interchangeable with factual acceptance or a universal rollback guarantee (MEM-RTE-review-decisions, EPI-RTE-refinement-decision, EPI-RTE-lint-assessment, EPI-RTE-repair-admission, EPI-RTE-repair-verification).
Delivery and continuation
Later recall resolves scope, searches and optionally reranks, then projects selected results. Standard MCP omits separate source text, structured fields and summary; a retained reason or generated recap summary therefore need not reach that client. Optional deep retrieval adds episode, fact-child and rollup channels; their existence does not establish default enablement. Host relevance judgments are instructed rather than enforced. Requested retrieval differs from automatic background selection into later inference tasks; the inspected session-start hook does not inject retained knowledge (RT-RTE-recall-search, RT-BAP-recall-skill, MEM-RTE-summary-rollup, MEM-RTE-brief-handoff).
Current-work continuation has a separate route. The shipped handoff procedure asks the host agent to derive typed Space Brief deltas from session evidence without ordinary per-item user approval. Later Brief reads reload the database snapshot and exact item identities, not a ranked fragment set. Its Markdown projection is a one-way receipt, not authority. Online updates use operation identities and version checks; offline replay can reset a summary's expected version, weakening that stale-summary boundary. The chronological session log is not established as an automatic later input (MEM-OBJ-space-brief, MEM-RTE-brief-handoff).
The representative connected workflow is an explicit single-page rebuild of a user-edited, non-authored page. The host skill requests destructive-action confirmation, MCP forwards target/force, and the daemon selects the first available foreground provider before delegating to shared refresh. No provider yields a skipped result; citation rejection or stale fencing can prevent a write. The foreground envelope has no retry loop: the caller owns a fresh attempt. Delivery is a receipt, not page prose. Its updated flag preserves the write outcome but omits gate, card and acknowledgement details. Consequently ok, updated: false cannot distinguish several no-write/staged outcomes. Another page or recall request is needed to obtain the content. This is an implemented conditional workflow, not an observed successful rebuild or API-enforced consent (WF-BAP-rebuild-skill, WF-RTE-rebuild-request).
What the architecture establishes
For a local capture-and-later-recall scenario, Origin supplies durable admission and later-consumer wiring. For current-project continuation, it supplies a typed snapshot path rather than search alone. For unattended page maintenance, it supplies candidate generation, selective ownership protection and a separate support/exposure architecture, conditional on configuration, eligible models and cutover. These conclusions concern mechanisms, not retrieval quality, factual reliability or improved answers.
The strongest supported learning contribution is automatic transformation of conversational experience into durable retrieval artifacts, plus an instructed session-to-Brief path consumed on later resume. This supports trace-fed artifact updating without weight training. It does not yet establish improved capacity for future action, or that criticism of a consumed theory caused such improvement. The improvement example in the README is explicitly illustrative, not an observed comparison (MEM-RTE-enrichment-slices, MEM-RTE-brief-handoff, EPI-CLM-notes-change, profile verification).
The independent properties remain separate:
| Property | Supported finding and unresolved question |
|---|---|
| Theory-builder condition 1: localized theory content | Not established for an actual theory instance. Storage and prompts can carry prose, but the records supply no specific theory-bearing unit to assess. |
| Condition 2: consumption of what the theory says | Context delivery is wired; theory-dependent decisions are not established by delivery alone. |
| Condition 3: content-directed criticism | Support judgments and contradiction handling exist, but no supplied episode establishes stated criticism, blame placement and resulting revision or changed reliance for a theory. |
| Condition 4: iteration | Retained revisions and later consumers exist; a criticism result shaping a next theory round is not established. |
| Reflection | Diagnostic and support state are inspectable, but the accepted account does not establish the required self-representation and both causal directions for a specified system aspect. No system-complete negative follows. |
| Autonomy | Scheduling and support publication have computational owners. The selected rebuild and handoff also depend on external host/human roles; an autonomous theory-builder loop is not established. |
| Self-improvement | Changeable retained context and prompt overrides provide behavior-shaping surfaces. The records do not establish an improvement-objective-driven update to Origin's own organization followed by operation causally dependent on it, nor a fully specified dormant pathway meeting that objective. Exercised self-improvement and successful improvement remain unestablished. |
These bounds follow the epistemic conclusion, RT-RTE-prompt-override and the trace-fed routes above. They do not collapse into one verdict. No comparison of improved future action was performed. Supplied before/after operation linked to retained changes would alter the benefit assessment; claim/evidence judgments with independent labels would alter the semantic-accuracy assessment. A traced theory/criticism/revision/next-round episode would permit condition-specific reassessment. Fuller writer and consumer tracing would change coverage, while uniform incarnation fencing or richer rebuild receipts would change the specific concurrency and delivery limits. Existing lab retrieval references alone cannot establish those outcomes or automatic production adaptation (EPI-RTE-retrieval-eval).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Grouped profile units combine distinct payloads and consumers. | SRC-1; profile-verification limit; MEM-OBJ-revision-cards, EPI-OBJ-lint-report, EPI-OBJ-repair-manifest, EPI-RTE-claim-judgment, EPI-RTE-repair-admission | Partial revision/lint/repair form unit and claim/support versus diagnostic/repair read-back unit. | Per-payload form completeness and one-to-one mechanism-level direction comparisons; supported existence findings remain valid. | Separate natural units with complete operative payloads and request/selection/delivery chains. |
| Implementation is not observed activation or causal benefit. | SRC-1; boundary, RT-RTE-ambient-work, EPI-RTE-retrieval-eval | Frozen source; no target execution, deployed databases, provider outputs or private benchmark caches. | Actual reliability, crash durability, improved future action, component effects and exercised self-improvement. | Inspectable configured runs, retained-result-to-consumer traces and appropriate controlled comparisons. |
| Page-support semantics and exposure depend on model, evidence and regime. | SRC-1; EPI-RTE-claim-judgment, EPI-RTE-support-publish, EPI-RTE-truth-exposure | Selected linked evidence, eligible judge, current-run predicate and conditional grants/generation. | Source truth, semantic soundness, exhaustive checking, actual accepted ampliative knowledge or universal supported-only delivery. | Actual content/judgments, independent semantic evaluation, provenance audits and deployed policy/grant state. |
| Included routes and payloads remain partly inspected. | SRC-1; MEM-RTE-import-inputs, EPI-RTE-refinement-decision, EPI-RTE-repair-admission, RT-RTE-remote-query | Graph/community/vocabulary, full import distillation and formats, every repair/refinement writer, UI/client projections, generated-skill parity, relay helpers and provider internals. | Complete admission, lineage, authority, permissions, semantic-withdrawal and client-parity inventories; complete architectural negatives on theory building or adaptation. | Targeted producer/payload/gate/consumer tracing for each unresolved included branch and external deployment evidence where required. |
| Refresh and foreground delivery guarantees differ by branch. | SRC-1; RT-RTE-page-refresh, WF-RTE-rebuild-request | Forced refresh without candidate sources versus ordinary/candidate-source writes; reduced foreground receipt. | Universal recreation protection, inference of a specific no-write cause, or page-body/review-card delivery from the receipt. | Branch-complete fence enforcement and a response preserving distinct outcomes and actionable identifiers, followed by content delivery evidence. |
| Usage feedback, compression and complete semantic withdrawal are not established across production paths. | SRC-1; MEM-OBJ-access-state, MEM-RTE-withdrawal-cleanup, MEM-RTE-eval-compression | Bounded count-flush and compression caller searches; selected dependency cleanup. | Attribution of ordinary recall-driven adaptive retention, production compaction, or retraction from every derived artifact/client copy. | Production caller chains, durable compression reload evidence and complete dependency/consumer invalidation traces. |
The accepted report verification declares no additional limits; no unresolved cross-report conflict was identified in the supplied reconciliation. The profile grouping limit above is carried without strengthening its partial units.