Nemsy synthesis

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

Bounded synthesis

From local material to a retained Wiki

Nemsy is a caller-directed personal knowledge assistant. It turns local documents and selected conversations into Markdown Wiki pages, then supplies retained text to later model calls. This account covers the shipped Python implementation, CLI, Web backend, browser callers, prompts, storage and evaluation definitions at commit 2338bab0aab96a91efb90a80249f33510b630260, with cutoff 2026-11-09. SRC-1 supplies implementation and doctrine/design evidence, not observed runs or causal experiments. The whole-system functional boundary excludes developer-assistant rules, roadmap claims and external dependency internals; no actual user Vault or service responses were supplied. See the boundary and evidence and Source register. This synthesis uses accepted members without additional source inspection.

The README's continuous-accumulation and autonomous-induction claims have a concrete partial realization: caller-triggered generation, page retention and index appending. They do not establish automatic maintenance of arbitrary existing pages. Python owns file selection, model invocation, parsing and filesystem effects. The remote model proposes text, not executable next steps. The inspected calls grant text generation rather than model-selected tools or agent delegation. The configured DeepSeek aliases and endpoint are mutable, not exact pinned model weights; provider-side state and parameter changes remain uninspected (EPI-CLM-knowledge-accumulation, RT-CMP-deepseek, runtime account).

The representative workflow is a CLI ingest followed by a separate human-requested query. Ingest does not automatically initiate the query. The CLI collects sources and uses stripped-content hashes to skip unchanged input unless forced. For each pending file, it supplies source text, ancestor directory guidance, the Wiki index and ten recent other pages, or fifty with wide context, to generation. Directory batches run serially: an earlier generated page can enter a later file's context if selected. That is an implemented state dependency, not demonstrated beneficial integration (workflow account, RT-RTE-document-ingest).

The model returns requested SUMMARY and QUESTIONS sections. Python requires SUMMARY, performs text cleanup, writes a dated source page and appends an index entry. It does not independently check source fidelity, truth, schema conformity or citation support before writing. Cleanup can remove substantive text by its position under a matched heading; it is selection and reshaping, not semantic validation. A source page can combine the document, guidance and prior Wiki content, and its requested provenance fields are not independently authenticated (MEM-OBJ-source-page, EPI-RTE-summary-cleanup).

Delivery and completion are different. A missing SUMMARY yields no page, but the enclosing ingest still records the nonempty source hash, while the CLI wrapper reports a completed call. A later unchanged CLI ingest can therefore skip a source that produced no page. Force or changed input permits another attempt; neither restores an earlier version. Page, index and log writes are separate, not a transaction. These are source-predicted consequences, not observed incidents (RT-RTE-document-ingest, MEM-OBJ-ingest-log, workflow recovery).

Controls depend on the caller. CLI directory exclusions do not cover every direct-file branch; HTTP direct-file ingestion bypasses collection and does not consume its force field. Empty-file recording occurs before the CLI dry-run return. Thus neither directory exclusion nor dry-run is a universal no-write or no-ingestion guarantee (RT-RTE-document-ingest).

From retained pages to answers and further retention

A later ordinary query receives the full Wiki index plus up to thirty recent other Markdown pages. Selection uses modification time and file count, not the question. The loader does not follow index links to recover omitted bodies and provides no model-accessible fetch tool for missing pages. Complete selected files and the unbounded index are not an input-token budget. The model interprets this window under instructions to answer only from Wiki content, cite pages and abstain when information is missing. That is a model-interpreted policy, not a checked answer invariant; absence from the selected window cannot establish absence from the whole Wiki (RT-RTE-wiki-answer, MEM-OBJ-wiki-index, RT-CLM-wiki-grounding).

Chat has different context lifetimes. CLI chat loads twenty recent pages once at session start; ingest or save does not refresh that snapshot. Web chat reloads twenty pages per request and accepts caller-supplied history. Positive CLI max_turns discards old message pairs without a summary; the shipped zero setting leaves history untrimmed. Browser Wiki mode omits history from that request but preserves the visible array for later Chat mode. Leaving the Chat view can instead lose its unsaved local state. The searched product paths contain no implemented cross-session raw-history or summary-cache checkpoint, despite configured directory names (RT-RTE-conversation-chat, memory annotations, MEM-ABS-history-store).

Retained material has several distinct effects. Wiki prose supplies model context; ancestor guidance supplies ingestion directions. Stored hashes enforce ordinary CLI skipping, while modification times rank pages for context. Token and operation logs supply accounting or action descriptions, not acceptance of propositions. Putting Wiki prose in a chat system prefix does not enforce a distinction between reliable knowledge and embedded instructions (RT-BAP-wiki-context, MEM-OBJ-index-guidance, MEM-OBJ-ingest-log, MEM-OBJ-token-log, MEM-OBJ-operation-log).

Optional ordinary query archive writes the answer with fixed metadata and removes its ARCHIVABLE marker. The caller's archive flag authorizes retention; the marker alone does not. HTTP query accepts but ignores archive. CLI reasoning query has a separate reachable tuple-to-string-consumer mismatch: conditional on provider completion, reasoning-plus-archive is predicted to fail when the archive helper calls replace on the tuple. Successful ordinary archive semantics therefore cannot be generalized to these alternatives (RT-RTE-answer-archive, MEM-OBJ-query-page, RT-RTE-reasoning-query).

Conversation saving is Nemsy's strongest supported trace-fed contribution. A CLI save or browser archive operation supplies selected user/assistant history to a fresh model request; code admits the rewritten insight page without review of its particular propositions. The page can subsequently reach generation through the Wiki loader. This is a wired conversation-to-durable-context pathway, independent of weight training, novelty or demonstrated improvement. It is not automatic history compaction: raw conversation is neither durably checkpointed nor replaced by the insight. Background and derivation are requested, not guaranteed, and save receives no freshly loaded Wiki despite requesting associated-knowledge links (RT-RTE-insight-save, MEM-OBJ-insight-page).

The accepted comparison distinguishes automatic Wiki provision from explicit file retrieval. Task-triggered code pushes a coarse recency-selected window without the model requesting those pages; a human choosing a file or citation pulls that retained material. Citation rendering and stem resolution make current page text inspectable but do not test whether it supports the answer. Duplicate stems, nonexistent references and changed page versions delimit that inspectability (MEM-RTE-browser-read, EPI-RTE-citation-display, profile verification). Ancestor-guidance selection is an additional route omitted from the comparison's separate direction/signal inventory; that inventory remains incomplete, as specified below.

Checking, maintenance and the assessment it supports

Nemsy has real but narrow checks. SUMMARY parsing gates format. Lint requests model diagnoses of contradictions, outdated views, orphans and gaps using index plus fifty recent pages, but returns advice without installing corrections or retaining a criticism-to-revision loop. Smoke definitions assert metadata, lengths, permissive answer substrings and archive-file existence; their result controls the smoke command's exit status, not normal knowledge admission. No passing execution was supplied, and test writes precede assertions without automatic rollback (RT-RTE-wiki-lint, EPI-RTE-smoke-check, EPI-RTE-smoke-disposition).

For the scenario of accumulating readable summaries and revisiting recent retained material, the implemented progression is established. For exhaustive corpus answering, verified knowledge production or automatic correction, it is insufficient: selection is windowed, source warrant is unknown, and generated text acquires operational use without evidence-consuming semantic acceptance. Actual source/output pairs would be needed to distinguish faithful compression, entailed derivation and conjecture; generation alone establishes none of these at proposition level. New pages and navigation links do not by themselves establish accepted knowledge or lifecycle integration (epistemic bounded conclusion).

The trace-to-insight route supports durable experience-derived context, but improved capacity for future action remains unestablished. Likewise, the inspected lint and retention routes do not establish content-directed criticism followed by retained revision shaping another round; this is not evidence of an autonomous theory builder. The runtime supplies no separate conditions 1–4 assessment from which to infer a stronger result. Reflection is not established merely because lint discusses the Wiki: no installed diagnosis-mediated update is shown. Exercised self-improvement is independently unestablished: later loading is wired, but no supplied operation demonstrates that an evidence-responsive organizational change was subsequently exercised with causal dependence, or that it improved an independently specified objective (MEM-OBJ-insight-page, RT-RTE-wiki-lint, epistemic bounded conclusion).

A traced save-to-later-use sequence could establish exercised retention and particular-content activation; controlled comparisons would be needed for benefit attribution. A semantic admission decision with criterion, evidence and intended scope would change the knowledge-acceptance assessment. An implemented consumer that applies, retains and reuses criticism would change the maintenance-loop assessment. None can be substituted by a confident answer, a successful counter or an archive label. The report verification accepts the bounded implementation account without asserting those stronger conclusions.

Limitations

Limitation Affected SRC IDs and record citations Inspected boundary Conclusion prevented Evidence that would resolve it
Ancestor-guidance delivery is omitted as a separate direction/signal comparison unit. SRC-1; MEM-OBJ-index-guidance, RT-RTE-document-ingest; declared profile limit. Extant nonempty ancestor files are selected farthest first and supplied to ingestion; the named Wiki-loader comparison unit does not cover that selector. No complete local selector inventory or all-guidance targeting comparison. The supported coarse Wiki provision and explicit browser retrieval remain valid; omission is not absence of guidance read-back or signal. Separate accepted classifications of this recorded selector, its trigger, inputs and delivered guidance.
No supplied operation or causal comparison. SRC-1; RT-RTE-document-ingest, RT-RTE-wiki-answer, RT-RTE-insight-save, EPI-RTE-smoke-check. Code and prompt/test definitions, not user Vaults, service traces or test results. No observed successful workflow, reliable grounding, particular-content activation, measured improvement or component-effect attribution. Inspectable end-to-end traces with retained inputs/results; appropriate interventions and comparisons for causal claims.
Unknown source warrant and generated-content relations. SRC-1; EPI-OBJ-source-content, MEM-OBJ-source-page, MEM-OBJ-insight-page. Import, generation, cleanup and syntactic admission; actual local provenance and source/output pairs unavailable. No established initial human authorship, fidelity, novel accepted claim or guaranteed preservation of reasons. Provenanced inputs, exact generated/retained outputs and evidence-consuming content assessments.
Context coverage and provider behavior remain bounded. SRC-1; RT-RTE-wiki-answer, MEM-OBJ-wiki-index, RT-CMP-deepseek. Local recency/count loader, full index and unpinned remote service aliases. No whole-corpus answer coverage, bounded input-token guarantee, effective cache benefit or comprehensive provider-memory account. Traced selections and token handling; provider contracts or inspectable state/version evidence for provider-specific conclusions.
Recovery does not synchronize withdrawal or restore versions. SRC-1; MEM-CLM-storage-lifecycle, MEM-OBJ-ingest-log, RT-RTE-conversation-chat. Separate writes, caller-trusted deletion helper, unchanged-hash skip and CLI snapshot. No atomic knowledge update, propagated retraction or automatic recovery from rejected summaries. Implemented and traced transactional/retry behavior, version restoration and derivative/snapshot invalidation.
Maintenance and improvement remain separate from retention. SRC-1; RT-RTE-wiki-lint, MEM-OBJ-insight-page, EPI-RTE-smoke-disposition. Diagnostic advice, durable insight eligibility and smoke exit-result consumption. No established iterative theory criticism, improved future capacity or exercised self-improvement. Recorded criticism, consequential revision and next-round use; causally linked subsequent operation and separate outcome evidence.
Deployment isolation and packaging are not demonstrated. SRC-1; runtime account, MEM-RTE-browser-read. Public bind, no endpoint authentication dependencies in inspected definitions, lexical containment, checkout frontend/config presence. No authenticated canonical filesystem sandbox, observed exploitation, functioning wheel deployment or source/bundle equivalence. Deployment and packaging evidence; authenticated access controls and canonical containment with their actual enforced scope.