Curated Thoughts synthesis

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

Bounded synthesis

Curated Thoughts is a file-to-knowledge system with model-assisted proposal generation, not a general-purpose agent executor. Its strongest supported contribution is an implemented path from ordinary source files to retained, source-linked facts that can inform later extraction and external agents. This establishes reusable context capacity, not demonstrated learning or improved task performance. Its stated ambition is that trustworthy short facts reduce the demands on weak consuming agents; the inspected checks establish narrower structural, lineage and selection properties, not that ambition's factual or performance result. See EPI-CLM-trustworthy-facts, MEM-OBJ-wiki-content and MEM-BAP-retained-fact.

The evidence is source inspection and declared intent in SRC-1, https://github.com/equationalapplications/curated-thoughts, at commit 433ebbd48c2ffd0e871e34a1505dcdef1a78c8fd, with cutoff 2026-11-08. The whole-system functional boundary includes the desktop, headless commands, both MCP servers, shipped prompts and skill, persistence, maintenance and retrieval. It is not an exhaustive behavioral audit. No target execution or user-vault/agent-use trace supports this account. External integration hooks, provider internals and the external wiki engine are outside the frozen implementation evidence; the local adapters remain included. See the boundary and evidence and Source register.

From a finding to a model proposal

The representative workflow is a task-solving agent depositing a finding after work and a later agent requesting relevant context before work. The shipped skill requests this ordering but does not enforce it. The source connects deposit, ingest, generation, admission and recall; it does not show a completed execution across those stages. The task-solving agent remains outside Curated Thoughts' inspected executor. See RT-CLM-agent-workflow and the workflow account.

A deposit supplies a vault-relative path, title and body. Rust validates the deposit lane and shape, exclusively creates a new file, records audit/kick state, and schedules ingest when generation is configured. MCP acknowledges a pending handoff; the CLI waits for its kick. Neither acknowledgement proves that a fact was produced. Lock contention can hand progress to a watcher. Audit failure attempts compensating file removal because file creation and SQL state are not one atomic transaction. See RT-RTE-deposit-kick.

Rust owns the next-step sequence. Folder mode, tier, note suppression and a content-hash/model-name watermark gate extraction. Seeded agent-lane rules permit synthesis with automatic approval; the fallback awaits review. The configured generation model proposes content rather than choosing arbitrary tools. Its request contains shipped guidance, selected retained entity summaries/fact bodies, numbered source chunks and structural context. Candidate selection uses name hits and mean chunk embeddings, with caps and clipping. The assembled user context is prefix-truncated to 4,000 characters, so candidates can survive while source or structural suffixes disappear. This is fresh bounded input assembly, not conversational compaction or a delegated agent with shared history. Outside material remains in files/SQLite and must be selected again. See RT-OBJ-folder-rules, RT-RTE-librarian-synthesis and the memory annotation.

Checking and admission do different jobs

Malformed JSON receives one corrective retry. Identifier filtering then uses the full pre-truncation collections, not just what the model saw. Unknown evidence labels can disappear while a nonempty fact proposal survives. That does not imply an unanchored fact reaches live storage: the later shared commit path separately skips a fact_add without a surviving chunk anchor. An anchor establishes lineage existence, not that the quoted source entails the fact. Imported source warrant and the generated proposition's semantic relation to its inputs remain unresolved; neither faithful reshaping, deduction nor new conjecture is established merely by the name “synthesis.” See EPI-RTE-evidence-filter, EPI-RTE-anchor-check and EPI-OBJ-proposal-items.

Admission may follow GUI accept/edit/reject decisions, CLI decisions or automatic accept-all policy. The resolver supplies transactional consistency, normalized-body deduplication, lineage persistence and an auto-approval-only deleted-source veto. Manual resolution is outside that particular veto; terminal presence does not authenticate a human. Committed facts, evidence and outbox requests are retention, not automatically an evidence-consuming acceptance decision for truth. Embedding repair can remain outstanding after commitment. See RT-RTE-proposal-admission and EPI-RTE-fact-retention.

Human confirmation is a distinct personal-vouch architecture: a decision stamps exact title/body content and changes human-tier delivery eligibility. Denial archives a fact and retains a document/chunk/text suppression key; undo can restore archived same-text matches. These are consequential trust and availability decisions, not measurements of truth or falsity. Later body editing and reconstruction do not continuously establish that every existing stamp endorses the current wording. See EPI-RTE-fact-confirmation, MEM-RTE-verification-state and MEM-RTE-human-maintenance.

Later consumption depends on the read surface

Retained prose lives alongside symbolic structure and trust metadata in SQLite; vault files and verification JSONL remain separate carriers. Dense embedding blobs affect ranking but do not establish changing model weights. Existing prose automatically reenters later librarian requests as context. External agents instead request selected results; returned facts advise rather than compel them. Proposal reasoning persists, but later candidate and direct-entry reads do not universally deliver those reasons. See MEM-OBJ-source-material, MEM-OBJ-wiki-content, MEM-OBJ-embedding-blobs and MEM-BAP-retained-fact.

The principal read paths must not be collapsed:

  • Contextual recall ranks wiki content lexically and code chunks by cosine, returning JSON rather than an answer. Full MCP requires a successful access-audit write; the separate tools MCP logs best effort. Its read-only knowledge surface does not mean a process making no writes. See RT-RTE-context-recall.
  • CLI wisdom matching applies scheme, eligibility, floor and two-stage selection. Chunk-hop selection can deliver a fact without its own fact-level cosine floor. Healthy activity ordinary-entry selection restricts provenance to human tiers, but message fallback and independent correction objects prevent making that restriction universal. Existing replacement chains can supply corrections even when ordinary entries abstain. See MEM-RTE-wisdom-match and EPI-RTE-match-selection.
  • Semantic graph/search and direct getters use other predicates. Their inspected filters exclude deleted rows but do not apply the CLI's supersession/expiry eligibility. A superseded fact is therefore not universally excluded from every read. See MEM-RTE-alternate-read.

Recall can also push a best-effort verification request to the desktop operator. A capped drain matches queued fact IDs to current database content rather than trusting saved display snapshots. This does not block recall pending human approval and is not automatic injection into an external agent. See MEM-RTE-verification-prompt.

Correction, recovery and assessment

Withdrawal includes explicit archive/denial, conservative source-grounding healing, delayed pruning and exact-source-reference hard erasure. Healing targets selected inferred rows and can refuse a bulk batch; missing or invalid grounding evidence can conservatively retain content. Pruning hard-deletes eligible inferred rows only after more than seven days archived. An outbox request does not prove remote deletion. Separately, a supersession deposit can remain pending: retaining a correction request and reading already-populated replacement chains do not close an automatic correction loop. See MEM-RTE-memory-withdrawal, EPI-RTE-source-forgetting and EPI-RTE-supersession-pending.

For a workflow needing retained, source-linked context and explicit review choices, the inspected architecture supplies those mechanisms. For one needing every returned proposition to be currently endorsed, every correction automatically applied, or every write governed by one gate, it does not establish those guarantees. Recovery remains distributed among callers, watchers, later dirty-document processing, pending review, transaction rollback and embedding repair. Logged generation failure can return a successful unit result without proposals, so an ingested ledger label is not an end-to-end completion criterion. See RT-RTE-librarian-synthesis and the workflow recovery account.

The bridge restriction is likewise local: the shared dispatcher rejects curated/wisdom tool families for its production bridge label, not all legacy reads or vault writes. No deployed OS confinement follows from that policy. External wiki-engine callbacks and alternative import/edit paths also prevent generalizing librarian validation to all memory updates. See RT-BAP-bridge-tool-gate and RT-RTE-wiki-maintenance.

The supported adaptive contribution is durable content reuse and mutable selection/trust state. Whether criticism of consumed guidance improves future capacity remains unestablished. Localized prompt guidance is inspectable; theory-content-dependent consumption, formulated criticism with blame placement, and criticism-driven iteration are not separately established as theory-builder conditions 2–4. Contradiction instructions alone do not establish their execution. Reflection, autonomous theory building and exercised self-improvement are also independently unestablished, not disproved. Automatic Rust scheduling is not evidence that every internal theory-building role is computational. No traced evidence-responsive organization change followed by causally dependent subsequent operation establishes exercised self-improvement here. Trace-fed learning is separately unresolved because deposit provenance does not establish original traces, and opaque alternative updates remain. See RT-OBJ-synthesis-prompt, MEM-OBJ-source-material and the epistemic conclusion.

A source-linked execution spanning deposited bytes, model-visible input, decisions, committed content and later consumer behavior would change the operational assessment. Semantic checks against sources and independently specified acceptance criteria would strengthen the warrant account. Controlled downstream comparisons would be needed for benefit attribution; independent component variation would be needed for component effects. The inspected frozen-vector evaluator instead checks a finite retrieval code path and was not executed in this run. See EPI-RTE-gate-regression.

Limitations

Limitation Affected SRC IDs and record citations Inspected boundary Conclusion prevented Evidence that would resolve it
Alternative admission and maintenance remain partially characterized; carried from report verification. SRC-1; MEM-RTE-human-maintenance, MEM-OBJ-wiki-content, RT-RTE-wiki-maintenance; verification limit Local wrappers and selected backends, not complete import conflict/trust, draft/merge, replay/repair, remote application or external-engine internals. Exhaustive admission/curation comparisons and universal provenance, rollback, denial suppression or durability. Positive local routes remain supported. Records tracing each alternative's actual retained parts, admission, transformation, recovery and later consumers; independently frozen dependency evidence where needed.
Current confirmation labels need not establish current-content endorsement; carried from report verification. SRC-1; MEM-RTE-human-maintenance, MEM-RTE-verification-state, MEM-OBJ-verification-ledgers, EPI-RTE-fact-confirmation; verification limit Exact-content confirmation operation, body edits and reconstruction without explicit clearing of every orphan stamp. Every presently human-labelled result was endorsed in its present form. Content-bound stamp invalidation/revalidation across edits and rebuilds, plus a trace tying the current body to its decision.
No supplied end-to-end execution or benefit comparison. SRC-1; RT-RTE-deposit-kick, RT-RTE-librarian-synthesis, RT-RTE-context-recall, EPI-RTE-gate-regression Source implementation, doctrine and finite evaluator design; no target execution here. Successful workflow completion, exercised acceptance, downstream activation, measured reliability or causal improvement. Linked execution artifacts and appropriately controlled outcome comparisons.
Structural checks do not establish semantic warrant or complete model visibility. SRC-1; EPI-RTE-evidence-filter, EPI-RTE-anchor-check, EPI-OBJ-proposal-items Prefix-truncated model input, full-collection validation and live-anchor commit check. Every admitted fact follows from a visible, warranted source; a determinate reshaping/derivation/conjecture classification. Captured inputs/outputs and proposition-level support and transformation assessment under named criteria.
Recall and capability controls differ by route. SRC-1; MEM-RTE-wisdom-match, EPI-RTE-match-selection, MEM-RTE-alternate-read, RT-BAP-bridge-tool-gate CLI core, fallback/corrections, lexical/direct/legacy reads and tagged dispatcher. Universal temporal exclusion, per-fact cosine floors, human-tier-only activity output or cloud isolation. Route-complete policy alignment and authenticated/deployed capability evidence, with alternatives explicitly tested or excluded.
Supersession requests do not establish application. SRC-1; EPI-RTE-supersession-pending, MEM-RTE-wisdom-match Pending-status branch and existing-chain correction reader. End-to-end automatic correction from deposited request to changed reliance. An implemented application/adjudication path and linked later-consumption evidence.
Trace origin, theory-building and improvement properties remain independently unresolved. SRC-1; MEM-OBJ-source-material, RT-OBJ-synthesis-prompt, RT-CMP-generation-provider, RT-RTE-wiki-maintenance Generic caller text, requested criticism, inference adapters and opaque providers/engine. A complete trace-learning classification, theory-building/reflective/autonomous verdict, or exercised self-improvement attribution; also system-wide negatives. Original update inputs and later consumers; records of content-directed criticism and iteration; self-representation feedback paths; objective-bearing update and subsequent causally dependent operation.