TheKnowledge synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From source intake to reusable pages
TheKnowledge is a personal knowledge gateway: it retains sources as local markdown and metadata, authors wiki pages, maintains them, and supplies selected content to later human and model questions. Python orchestration owns the next step; ordinary models return scores, update plans or answer prose rather than independently operating the vault. This is a knowledge and context-engineering system, not an enclosing general-purpose agent runtime. The analysis covers its shipped CLI, MCP and browser interfaces, background drivers, agent instructions, retained-state lifecycle and evaluators at Git commit aca6e29e69c6a72b697b8178841d03e6296ee3fc, with cutoff 2026-10-14. SRC-1 supplies implementation and doctrine/design evidence only. No target execution, tests, provider calls, observed runs or causal experiments establish activation, reliability or benefit. Whole-system functional inclusion does not mean every handler or retained claim was assessed. See the boundary and coverage and Source register.
The representative workflow is wiki ingest <input> --domain <domain> --with-plan for a registered domain. It connects source conversion, editorial filtering, retained-page selection, model-proposed full-page updates, validation, repair and CLI delivery in one shipped invocation. It is not a team of delegated agents. The filter proposes a score; code thresholds decide source-page inclusion. Raw source retention is independent: a review or rejected source can remain on disk without generating wiki pages. On the new-source branch, authorship additionally requires a written source page. Calling the command is not approval of each generated page. RT-RTE-ingest-authorship; selected workflow.
Authorship receives capped source text and selected existing pages, not the entire vault. Its default model call is tool-free and does not persist a model session. Code parses the returned plan, validates all updates and allows one error-fed repair attempt; the repair includes at most 20,000 characters of the previous response. A model-call exception fails immediately. The model proposes both prose and frontmatter: truthy returned draft metadata can downgrade missing-section and uncited-claim errors to warnings even when the caller omitted a draft flag. Other integrity checks remain active. Warnings on a successful application do not by themselves trigger repair. Thus successful application is not uniformly strict admission or finalization. RT-OBJ-plan, RT-BAP-authorship-prompt, workflow qualification.
Accepted updates use per-file atomic replacement with snapshots and best-effort exception rollback, not a crash-atomic multi-page transaction. Source admission survives authorship failure; CLI success can mean source-only success with a warning, or a no-op rather than newly authored pages. Local application is not a CommitGate Git commit. These distinctions define what delivery means before any later consumer receives the output. RT-RTE-ingest-authorship; workflow delivery.
Alternate admission and later reliance
Asynchronous deposits follow a different route. MCP returns a durable intent receipt; a separately running shipped commit-worker must claim, render and submit it to CommitGate before canonical writing. Rendering passes caller markdown verbatim. The gate supplies lease fencing, history-based redelivery checks, deduplication, routing and Git admission, but not the synchronous plan route's full page/citation validation. Its lock differs from the plan lock, and HEAD comparisons do not protect live uncommitted edits. Source deposits can pass enqueue shape checks but fail rendering; a volatile receipt's noncanonicalization flag is not enforced by the inspected ordinary drain. A receipt therefore establishes neither completed authorship nor every advertised downstream condition. RT-OBJ-intent, RT-RTE-deposit-commit.
Other gateway writers also differ: local section/concept edits validate through their own mechanisms; daily digests directly admit model-written drafts; artifact filing checks remote completion/download rather than semantic page grounding. Gateway-only writes are foreground instructions, not filesystem isolation or one universal quality gate. Normal plan/filter calls, vision calls with Read access and permission bypass, direct SDK answering, and scheduled shell commands do not share one execution envelope. HTTP authentication and an MCP read-tool allowlist protect their particular mounts, not every host capability. MEM-RTE-local-amendments, MEM-RTE-digest-authoring, MEM-RTE-artifact-filing, RT-CLM-gateway-only, runtime controls.
Later question answering selects retained sections through lexical ranking and optional dense fusion for a single domain; multi-domain retrieval uses lexical quotas and interleaving. SQLite indexes and numerical vectors are derived access state, separate from canonical page text, which is reread for delivery. Drafts are eligible by default and marked. Trust contributes a small page-level ranking term, not a per-claim truth judgment or exclusion. Long-section truncation can lose citations, and the first block may exceed the requested total budget. Whole-page neighborhood fetch is a separate requested read: its root remains whole and its JSON return ignores the budget. MEM-OBJ-search-index, MEM-OBJ-section-vectors, RT-RTE-retrieve-answer, MEM-RTE-neighborhood-fetch.
The answer model receives selected blocks and a context-only instruction. Cleanup deletes citation IDs absent from those blocks but leaves prose, including potentially unsupported prose, intact. Page citation checks likewise test citation-shaped coverage, not claim/source entailment. For ordinary cited assistance, these mechanisms provide inspectable context selection and attribution controls. For a scenario requiring factually accepted claims, strict context ceilings or complete retained provenance, those controls are insufficient evidence of the required guarantee. No inspected proposition/input comparison establishes checked derivation or accepted novel knowledge; model generation alone establishes neither. EPI-RTE-answer-cleanup, EPI-RTE-citation-check, EPI-CLM-retrieval-grounding, EPI-CLM-answer-grounding.
Remote research adds acquisition, filtering, notebook construction, source maps and synthesis filing. Optional plan review is not compulsory output approval. Failures through application abandon the session; persistent-corpus promotion can fail after local pages were written without revoking them. Notebook IDs, citation-diversity probes and opaque corpus tokens do not disclose the provider's full indexed or model-visible evidence. Artifact wrappers are also not decoded payloads: the briefing path overwrites its downloaded markdown report with its wrapper, and normal wrapper creation supplies an empty explicit source list. RT-RTE-research-query, EPI-CLM-citation-provenance, MEM-OBJ-artifact.
Maintenance, feedback and changes to future behavior
Maintenance changes availability and reliance without necessarily deciding truth. Strict finalization can promote draft status, but does not establish factual support. Contradiction scans propose diagnoses and can retain weekly draft reports for later retrieval. Deposit conflict selection instead uses source-category precedence and string comparison: incoming category is caller-derived, and equal-trust old/new labels favor the existing claim rather than measuring actual recency. Both claims remain; the Contested annotation is interpretive, not claim-specific ranking enforcement. Retraction/supersession flags signal review without excluding old claims in inspected retrieval. De-path cleanup can remove a page while preserving provenance-fed restoration; restoration reintroduces text, not warrant. RT-RTE-draft-finalization, EPI-RTE-contradiction-scan, EPI-RTE-conflict-selection, EPI-RTE-conflict-materialization, MEM-RTE-source-maintenance, MEM-RTE-corpus-recovery.
The strongest supported feedback contribution is durable reuse of corrected editorial decisions and reasons. Software pins examples; later filter calls select labels, scores and rationales as guidance for future scoring. This is a wired trace-fed context route, not demonstrated increased accuracy. Original caller identity and controlled trace-origin categories remain unresolved. The weekly diagnostic draft also supplies durable context from checking retained claims, although its route-specific memory classifications remain incomplete. MEM-RTE-example-correction, MEM-BAP-example-calibration, profile verification limit.
Example-fed distillation can write revised criterion candidates and optional label-agreement metrics, but stops before live promotion. Bootstrap can directly install schema-valid policies. Queued policy-edit has principal and evaluator gates, but its retrieval floor measures current FTS behavior and its proposed dedup check covers only that tested domain; these are not universal policy-quality guarantees. Wiki-state model judgments retain scores and trends for display rather than automatically revising pages. Generated domain skills derive policy, navigation links and synthesis titles, not the recent operational logs their description suggests; later host loading is afforded, not observed. RT-RTE-policy-distillation, EPI-RTE-filter-calibration, RT-RTE-policy-bootstrap, RT-RTE-policy-edit, MEM-RTE-evaluation-retention, RT-RTE-skill-emission.
These partial capacities leave independent properties unresolved, rather than forming one verdict:
- Theory building: localized policy content and content-dependent consumption satisfy conditions 1 and 2 at wired status for distillation/bootstrap; policy-edit affords condition 1 and wires condition 2. Conditions 3 and 4 remain uninspected: neither schema retries nor metric comparisons establish stated content-directed criticism with blame and retained results shaping another criticism round. RT-RTE-policy-distillation, RT-RTE-policy-bootstrap, RT-RTE-policy-edit.
- Learning from criticism: example reuse and candidate revision are implemented contributions. Whether criticism of consumed policy improved future capacity remains uninspected; no comparison supports that benefit or its attribution. The memory notion of a trace-fed write does not establish this stronger result. MEM-RTE-example-correction, RT-RTE-policy-distillation.
- Reflection: prior decisions can be represented and affect candidate generation, but the distillation account does not establish both live causal directions through installed policy. Bootstrap and change-control alone do not establish reflection. Autonomous theory building separately remains uninspected: computational authorship does not resolve criticism and iteration, and declared human promotion remains outside candidate generation. RT-RTE-policy-distillation, RT-RTE-policy-bootstrap, RT-RTE-policy-edit.
- Self-improvement: within editorial source selection across future requests, correction reuse offers a wired path for evidence-responsive changes to durable guidance. It does not establish exercised self-improvement: no subsequent operation is shown causally depending on the change, and no independently measured selection-quality gain is supplied. Distillation additionally leaves live installation unresolved. This is not evidence that all self-change is absent. MEM-RTE-example-correction, RT-RTE-policy-distillation.
Development-session continuation is weaker than automatic checkpointing: hooks emit instructions asking the foreground agent to write/commit selected working memory and later reread/verify it. Scripts perform neither action themselves; host compaction products and differing reread predicates remain unresolved. A saved path is a reload cue, not content automatically present in a model window. MEM-OBJ-session-state, RT-RTE-session-reanchor.
The assessment would change with linked execution evidence showing actual admission, retained outputs and later consumers; claim/source comparisons testing support after selection and truncation; live candidate-installation and subsequent decision traces; and controlled outcome comparisons for feedback benefit. Strict budget enforcement, explicit draft-mode control, shared writer coordination or withdrawal checks would change the corresponding architectural guarantees if implemented and traced. Until then, the supported result is reusable retained context and route-local control, not universal grounding, exhaustive maintenance coverage or demonstrated improvement.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied execution or causal comparison | SRC-1; RT-RTE-ingest-authorship, RT-RTE-retrieve-answer, MEM-RTE-example-correction | Frozen implementation and doctrine, not target runs | Exercised admission, reliable delivery, factual quality, beneficial learning and causal improvement | Provenance-linked runs through later consumers and objective-specific comparisons |
| Policy/source-curation inventory is incomplete; report verification carries this limit | SRC-1; RT-RTE-policy-edit, RT-RTE-policy-bootstrap, memory inventory, report limit | Selected routes; verification additionally identifies direct proposal promotion, legacy-policy conversion and inbox-triage metadata writes, without complete retained chains | Exhaustive policy admission, rollback and source-curation route comparisons | Retained route accounts for promotion, legacy conversion and triage, including their later consumers |
| Vector invalidation and reverse-merge preservation remain partially inspected; report verification carries this limit | SRC-1; MEM-OBJ-section-vectors, MEM-RTE-corpus-recovery, EPI-RTE-depath-recovery, report limit | Cache mtime/dimension checks and provenance-fed public restoration; incomplete update/removal and reversal chains | Universal ranking freshness, exact encoder compatibility and preservation of intervening edits | Complete edit/delete/rebuild and reverse-merge traces, including overlap and concurrency cases |
| Diagnostic route classifications are incomplete; profile verification carries this limit | SRC-1; EPI-RTE-contradiction-scan, MEM-OBJ-wiki, MEM-RTE-example-correction, profile limit | Automatic weekly diagnostic draft and later retrieval are established, but grouped with unresolved histories in the profile | Complete diagnostic route counts and route-by-route authority, delivery and original trace-source comparison; not the independently wired existence of trace-fed guidance | Separate retained classifications of diagnostic production, delivery, consumers and original input content |
| Other included production and maintenance paths are not fully assessed | SRC-1; MEM-OBJ-raw, MEM-RTE-local-amendments, MEM-RTE-artifact-filing, memory inventory | Converter/poller/migration lineage, additional editing, domain lifecycle, synchronization, exports and maintenance remain locally incomplete | Complete memory inventory, universal admission guarantees, or system-complete negative about knowledge production | Route-specific inspection and retained records for consequential alternatives |
| Providers, initial provenance and host context remain opaque or uninspected | SRC-1; RT-CMP-notebooklm, RT-CMP-qwen3-embedding, MEM-OBJ-examples, MEM-OBJ-session-state, RT-RTE-session-reanchor | Local request identifiers, wrappers, examples and hook instructions, not exact weights, remote selected corpus or host compaction products | Reproducible provider reasoning, full lineage, controlled original trace-source categories, or guaranteed checkpoint creation/resumption | Versioned provider/corpus evidence, original feedback provenance and host traces linking snapshot inputs, writes and rereads |
| Structural checks and surviving citations do not establish semantic warrant | SRC-1; EPI-RTE-citation-check, EPI-RTE-answer-cleanup, RT-RTE-research-query | Citation-presence/membership checks and opaque remote mapping; individual retained claims unassessed | Claim truth, checked derivation, accepted novel knowledge or complete provenance after truncation | Actual proposition/input comparisons and evidence-consuming acceptance decisions with criterion, intended use and scope |
| Evaluator references and licenses are bounded | SRC-1; EPI-RTE-wiki-evaluation, EPI-RTE-retrieval-measure, EPI-RTE-filter-calibration, RT-RTE-policy-edit | Selected context, expected-slug ranking, binary labels and proposed dedup behavior; golden provenance unresolved | General answer correctness, independence of calibration labels, all-field policy quality or demonstrated benefit | Validated references, documented independence, broader target-specific tests and observed results |
| Withdrawal and recovery are consumer- and timing-specific | SRC-1; MEM-RTE-source-maintenance, EPI-RTE-resolution-revert, EPI-RTE-orphan-selection, RT-RTE-deposit-commit | Review flags, coarse annotation matching, scan-time reachability and cooperating gate workers | Universal local/remote withdrawal, safe delayed cleanup, all-writer serialization or crash/power-loss guarantees | End-to-end consumer withdrawal and concurrency/recovery evidence, or enforcing changes covering the alternate routes |
| Revision paths do not establish criticism-driven improvement or autonomous theory building | SRC-1; RT-RTE-policy-distillation, RT-RTE-policy-bootstrap, RT-RTE-policy-edit | Localized criteria, candidate generation and route-local admission; criticism, iteration and live deployment unresolved | Theory-builder conditions 3–4, reflective/autonomous theory building, and exercised self-improvement across future requests | Stated criticism with claim/result/blame, retained next-round use, deployment and subsequent causally dependent operation; separate outcome comparison for benefit |
No unresolved cross-report conflict was retained in the supplied reconciliation. No frozen target source was read for this synthesis; supporting passages remain in the cited members.