Link synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From source material to retained context
Link is a local memory and wiki layer for external agents, not an enclosing agent runtime. It retains personal/project claims, source-backed articles, session proposals and task handoffs, then supplies selected context to later sessions. This account covers its shipped CLI, MCP, HTTP viewer/mutations, LinkBar callers, instructions, maintenance and evaluation surfaces at commit cc92f3632121b04063b24eaaf3a8c5429eaf2b57, with an applicability cutoff of 2026-10-14. The evidence is frozen repository inspection: implementation, declared procedures and attributed benchmark reporting. No target execution, observed host session or causal experiment is supplied. External agent hosts, their models and actual user judgments remain outside the inspected system; Link does not own their final answers or application actions. See boundary and its Source register, SRC-1.
The durable content is editable Markdown. Claim text and procedures reach interpreting agents; lifecycle/project fields reach fixed code selectors. Graph/search caches, SQLite full-text search and optional dense vectors provide access state rather than an independently authored knowledge corpus. Vector refresh is not provider-weight training, and the external embedding and reranker models are not revision-pinned. See MEM-OBJ-memory-claim, MEM-OBJ-memory-controls, MEM-OBJ-wiki-cache, MEM-OBJ-semantic-index and RT-CMP-reranker.
Explicit remembering passes supplied text through heuristic secret, duplicate and conflict checks. Their disposition can refuse a write, but caller overrides alter that policy. A successful write is active and pending review: pending is not a quarantine excluding it from recall. The requirement that a human ask or approve is a shipped agent protocol, not an independently authenticated approval credential checked by the core. Review subsequently changes ranking, warnings and due dates; its optional note and callable marker do not establish a particular evidence-consuming accuracy judgment. See RT-RTE-memory-admission, EPI-RTE-memory-check, EPI-RTE-memory-disposition and EPI-RTE-review-marking.
Session-end capture is a different stage. Automatic hooks mine bounded user turns, suppress echoes, duplicates and dismissed proposals, and retain proposal-only notes with review context. Acceptance later re-mines those notes and calls the memory writer; it does not automatically approve memory or consume an immutable numbered proposal checkpoint. Manual notes and imported instructions have different speaker provenance. Normalization can remove speaker prefixes or temporal qualifications, so general semantic preservation remains unresolved even though automatic user-only extraction narrows the risk. See RT-RTE-session-capture, MEM-OBJ-session-capture and EPI-RTE-proposal-normalization.
Authoring, checking and delivery
The representative multi-step workflow is human-requested ingestion of an arbitrary raw source. Link's status/skill guidance identifies pending files and withholds its suggested prompt on detected secrets or access failure. The foreground external agent then reads the source, discusses takeaways, creates or updates source/concept pages, records contradictions, rebuilds indexes and validates before reporting completion. That agent owns interpretation, writes, repair and the final report; no second worker or internal autonomous compiler is established. A source-processing request authorizes the workflow, not separate assent to every wiki claim. This is a shipped procedure, not a recorded successful episode. See workflow account, RT-RTE-agent-wiki and WF-RTE-wiki-validation.
Validation checks finite structural, link, date and secret-pattern conditions. Default validation can pass with warnings. In agent authoring it runs after persistence and constrains completion reporting by instruction; failure does not automatically undo files. Structured export ingestion instead checks ownership hashes, stages generated pages and enforces validation before live copying. These are different enforcement points, neither a truth oracle nor a universal pre-write wiki gate. Structured rendering preserves documented source lineage with an explicit bracket-escaping exception; its counts concern supplied export records and its high-confidence label is preset. Source links, author confidence tags and successful structure checks do not establish source truth or entailment of authored prose. See EPI-RTE-wiki-validation, EPI-RTE-staged-disposition, EPI-RTE-export-rendering and EPI-CLM-confidence-tags.
An ordinary substantive MCP recall invokes deterministic query construction: focused memory matching, generic wiki search and graph neighbors become a bounded packet and shorter capsule. The Python executor returns context; the host decides whether and how to use it. Optional embeddings affect memory similarity, while explicit memory-only reranking is separate from the ordinary combined packet. Larger budgets, full pages and direct files remain available on demand. Link keeps source bodies out of initial supply; it does not evict prior host turns, create a separate agent frame or enforce a host token-window cap. Fixed clipping can omit qualifications, and capsules need not retain all richer warning or procedure fields. See RT-RTE-context-recall and EPI-RTE-context-compaction.
The decisive selection distinction is between focused memory and generic wiki delivery. Project/lifecycle predicates gate focused recall and the combined packet's memory branch. Generic wiki search instead considers all nonhidden Markdown, including archived, expired or other-project memories, and can return them as pages without memory-specific review/lifecycle warnings. Actual delivery depends on query, corpus and ranking; no disclosure was observed. Archive therefore withdraws focused default eligibility, not all-packet availability. Updating can also change matching through appended body text while compact delivery still shows the old TLDR. See MEM-RTE-wiki-read, EPI-RTE-memory-withdrawal and MEM-RTE-memory-lifecycle.
Context is not exclusively requested. Startup hooks and the first suitable MCP response independently supply briefs and pending handoffs. Handoffs relay priority task text with expiry, not a live worker or the previous hidden reasoning state. A prompt guard selects a strongly overlapping saved constraint and asks the agent to confirm a conflict; it does not mechanically block an action. Its retained usage history can suppress repeated reminders. Server-process autobrief state is not a proven host-session identity, and MCP guard project forwarding differs from the prompt hook. See RT-RTE-session-delivery, MEM-OBJ-handoff, MEM-RTE-guard-reminder and report verification limits.
Later change and what it establishes
The strongest supported learning-related contribution is automatic durable trace-fed context/control: session material becomes later review/startup context; recorded reminder events shape cooldown; operation history shapes later recovery. These are wired later-consumer paths, not merely raw logging, and qualify as trace-fed writes in the memory comparison's sense. They do not establish improved capacity for future action. Usage events measure software surfacing, not model reading or benefit; popularity ranking remains an optional library capability rather than an adopted ordinary-adapter feedback loop. The consolidation command proposes actions without installing a semantic successor. See MEM-RTE-usage-feedback, RT-OBJ-operation-journal, RT-RTE-local-recovery, MEM-RTE-consolidation-plan and profile verification.
The runtime's theory-builder assessment remains separate for each condition: localized theory content is afforded by page templates, with no actual candidate supplied; theory-dependent consumption is uninspected; content-directed criticism is claimed by contradiction-checking instructions, without observed blame/test/revision decisions; criticism-shaped iteration is uninspected. No evidence establishes that criticism of a consumed theory improved future capacity. Lint affords representation and editing of wiki aspects, but an exercised reflective loop is unestablished. Autonomous theory building is also unestablished: human choices and opaque host interpretation remain material. These are independent findings, not a bundled negative. See RT-RTE-agent-wiki.
Self-improvement at this local-system boundary is likewise unestablished. Durable updates and later behavioral channels are supported, but no supplied episode traces objective-relevant evidence through determination of an improvement-directed organizational update to subsequent operation causally depending on it. Automatic feedback or changed files alone do not close that chain. The offline recall evaluator compares authored target ranks and can fail its process on regression; hygiene/poisoning evaluators test synthetic governance conditions. Neither installs an automatic product successor, and neither was run here. External reported QA results lack their separately retained adapter/judgments. Thus no demonstrated learning benefit or component-causal effectiveness follows. See RT-RTE-recall-evaluation, MEM-RTE-memory-evaluation and report verification limits.
For an explicit local-memory scenario, the evidence supports inspectable persistence, route-local admission controls and compact later retrieval. For a scenario requiring verified approval, universal withdrawal or warranted factual answers, those stronger conclusions do not follow. Git/team import, direct edits and restoration admit existing bytes through different gates; backups, exports and caches also prevent a blanket all-copy forgetting claim. Journal snapshots and per-file replacement support best-effort recovery, not workspace-wide serialization. See RT-RTE-workspace-sync, RT-RTE-local-recovery and MEM-RTE-retained-export.
The assessment would change with traced approval and host-consumption episodes, source-to-claim judgments, and controlled later-capacity comparisons at a named horizon. Cross-branch eligibility enforcement would change withdrawal guarantees; refreshing summaries on update and preserving qualifications through compaction would change the visible-content assessment. Pinned provider weights and inspectable QA artifacts would enable more specific evaluation. A theory candidate with consumed criticism and a criticism-shaped next round would change the corresponding theory conditions, without automatically proving reflection, autonomy or self-improvement. No source files were read for this synthesis; its evidence remains the accepted member records.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Changed bodies, temporal eligibility and compact excerpts differ. | SRC-1; MEM-RTE-memory-lifecycle, EPI-RTE-memory-update, EPI-RTE-context-compaction; report limits | Append/reset-review, temporal selector and packet/capsule construction | Updating necessarily refreshes the visible claim; temporal recall restores historical bodies; provenance preserves full semantic scope | Summary-refresh and historical-version mechanisms; source-to-delivery fidelity checks including qualifications and richer fields |
| Reminder project targeting differs by surface. | SRC-1; MEM-RTE-guard-reminder, RT-RTE-session-delivery; report limits | MCP guard call omits resolved project; prompt hook supplies it | Cross-surface project-targeting equivalence or reminder access control | Consistent forwarding and branch-specific selector/delivery evidence |
| Maintenance/authoring inventory omits detailed software shape repair. | SRC-1; RT-RTE-agent-wiki, EPI-RTE-wiki-validation, WF-RTE-wiki-validation; report limits | Verifier inspected doctor title-derived summary/raw-reference insertion and file writes | Exhaustive authoring/repair-route comparison; structural repair as faithful factual synthesis or truth acceptance | Accepted records characterizing doctor repair inputs, outputs and consumers, with semantic evidence for any stronger claim |
| No deployed operation or causal effectiveness is supplied; evaluator coverage and model identity remain bounded. | SRC-1; RT-RTE-recall-evaluation, MEM-RTE-memory-evaluation, RT-CMP-model2vec, RT-CMP-fastembed, RT-CMP-reranker, RT-BAP-context-packet; report limits | Inspected implementation and synthetic evaluators; external weights/QA judgments unavailable; other accessible evaluators not comprehensively assessed | Empirical approval/compliance, factual answer quality, learning benefit, component effects, exact model version or complete evaluator comparison | Installed-host traces, frozen weights, QA adapter/judgments, remaining evaluator analysis and controlled comparisons |
| Raw-source object inventory is coarser than import/read-route coverage. | SRC-1; RT-RTE-source-import, RT-RTE-project-seed; profile limits | Raw Obsidian copies and project-seed provenance snapshots are not separately classified across object axes | Complete raw-source inventory and per-object provenance-consumer comparisons; not the positive finding of raw retention | Separate accepted object/consumer records and corresponding partial-inventory classifications |
| Digest, auxiliary payloads and some presentation/admission details remain locally uncharacterized. | SRC-1; MEM-RTE-consolidation-plan, MEM-OBJ-wiki-pages, MEM-RTE-retained-export, RT-RTE-context-recall; memory coverage | Available included source; partial transformation, caller-control, rendering and client inventory | Complete memory-axis inventory, client parity, semantic-preservation or system-complete absence claims | Accepted records covering unresolved mechanisms, actual consumed payloads, admission control and initial provenance |
| Approval/review and source warrant are not authenticated by retention or labels. | SRC-1; RT-RTE-memory-admission, EPI-RTE-review-marking, EPI-CLM-confidence-tags, EPI-RTE-proposal-normalization | Overrideable writer, optional review note, authored tags and mixed manual/import provenance | Universal human assent, accepted claim-instance accuracy, calibrated confidence or general normalization fidelity | Content-specific approval/evidence records, checked source interpretations and speaker/time-preserving transformation evidence |
| Eligibility and deletion are consumer/copy-specific. | SRC-1; MEM-RTE-wiki-read, EPI-RTE-memory-withdrawal, MEM-RTE-retained-export, RT-RTE-local-recovery, RT-RTE-workspace-sync | Focused filters, independent generic pages, retained copies and per-file recovery | Universal archive/project nondelivery, all-copy erasure or whole-workspace transactional recovery | All-consumer eligibility/copy handling and explicit erasure/transaction guarantees with covered alternate paths |
| Theory use, criticism-led benefit and independent system properties lack episode evidence. | SRC-1; RT-RTE-agent-wiki, MEM-RTE-usage-feedback, RT-RTE-recall-evaluation | Templates, criticism instructions, wired metadata feedback and offline comparisons | Exercised theory consumption/iteration, criticism-improved capacity, reflective/autonomous theory building or self-improvement attribution | Candidate/decision/criticism traces with blame and retained next-round effects; property-specific causal links and later-capacity comparisons |