Atomic synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Atomic is a personal knowledge system with model-dependent enrichment, chat, wiki synthesis and recurring research reports. This account covers the shipped repository at commit 983bea6baf449d060514c94b3b15b5f85f77ba9b, with an applicability cutoff of 2026-10-12. SRC-1 supplies inspected implementation and attributed design claims. No target execution, provider trace or experimental result is registered. The whole-system boundary includes standalone and cloud drivers, clients, imports, maintenance and evaluators; it does not imply that every branch was audited. External model services and MCP clients' enclosing runtimes are outside Atomic's inspected responsibilities. See boundary and evidence.
Capture becomes selectable state
Atomic stores Markdown notes as atoms. Core creation inserts content before requesting embedding and tagging; command-line import can instead insert pending rows directly. A returned atom therefore does not prove completed enrichment. Chat can also create or edit atoms: the model proposes content, while compiled handlers check shape and concurrency and perform host-storage effects. Their instruction to write only on explicit request is model guidance, not an independent consent gate. Scoped readers do not establish a universal write boundary: the inspected edit handler loads its supplied ID directly. See RT-RTE-capture-write, RT-RTE-atom-mutation and RT-CLM-explicit-mutation.
Retained content and access state have different roles. Text chunks supply selected prose; dense vectors influence retrieval; tag membership and hierarchy determine scope eligibility. Embeddings are not provider weights. Single-atom enrichment automatically replaces chunks and applies generated tags, subject to configuration and availability checks. Replacement deletes previous chunks before successful embedding is assured, and tagging receives a limited text prefix. Standalone schedulers and cloud dispatchers maintain this state. A graph batch can be marked complete after a caught computation error, so completion status is not a semantic-success guarantee. See MEM-OBJ-atom-chunks, MEM-OBJ-chunk-embeddings, MEM-OBJ-tags, MEM-RTE-atom-enrichment and RT-RTE-state-maintenance.
Tag compaction is a separate taxonomy operation, not conversational summarization. A model proposes merges; fixed hierarchy guards constrain their application. In the inspected SQLite path, the winner receives associations and children, and the loser is deleted. Reasons are logged rather than established as durable rationale read by later taxonomy consumers. Explicit caller-supplied pairs bypass model proposal selection. These are implemented changes to future selection, not evidence of better categorization. See MEM-RTE-tag-compaction.
A bounded agent consumes the store
For a chat turn, Atomic saves the user message, reloads conversation history, assembles system guidance and context, and invokes a configured provider. The model chooses answers and tool requests; Rust owns registered-tool lookup, sequential execution and persistence. The ordinary loop has ten model rounds, not a ten-tool bound; a tool-free salvage call can follow the cap. Unknown names fail local dispatch. Cancellation is cooperative and does not interrupt an already awaited tool or undo completed writes. Assistant text, tool calls and citations are saved in successive fallible steps. These controls belong to this route, not every model call or external host. See RT-RTE-chat-turn and RT-OBJ-tool-registry.
History supplies prior role/content text automatically on a new turn, but historical tool-call and citation metadata are not reconstructed for the provider. Context trimming drops older groups without generating a durable summary. Mandatory first/latest groups can exceed the nominal budget. Requested corpus search and ID readers form separate read-back paths. Chat search admits all atom kinds by default; MCP search defaults to captured content. Agentic wiki research also selects captured content. This distinction excludes report-kind findings on those paths, not all generated prose: saving a chat answer converts it into an ordinary captured atom, with reduced wiki/finding citation lineage. See MEM-OBJ-chat-history, MEM-RTE-corpus-retrieval and MEM-RTE-answer-save.
Research reaches a retained product
The representative workflow is running an existing report now and retrieving its finding. It is a connected source-supported lifecycle, not an observed successful episode. HTTP returns 202 dispatched and a findings URL before research completes. The runner claims durable work, resolves source scope/window/cap, seeds citation identities and supplies source snippets. The model uses a bounded research loop with read, search and done tools. A separate final inference call receives the research transcript and writes prose; it is not an independent factual review. See workflow account and RT-RTE-report-finding.
The final parser checks response shape, including a citation trailer and nonempty body, and permits bounded retries or a corrective nudge. Persisted citations come from body markers resolved against the ledger, not the trailer's list. The runner automatically admits nonempty final prose. SQLite writes the finding atom, tags, provenance and citations transactionally, but cache and ledger completion follow separately. Execute-stage failures can defer, back off or abandon; success-branch bookkeeping can fail after the finding exists. Empty results advance the source-window watermark without a finding, and capped batches need not cover every eligible atom. Delivery means a stored, retrievable finding with best-effort notification—not dispatch acknowledgment, client receipt or exactly-once completion. See workflow account, MEM-OBJ-report-findings and EPI-RTE-citation-resolution.
Wiki publication offers a different admission choice. Pending proposals can be retrieved, accepted or dismissed; acceptance checks their base article identity and timestamp. Immediate propose-and-accept and direct generation bypass separate review. Prior SQLite article prose and citations are archived, but historical readers do not establish automatic restoration. Live, pending and archived products consequently have distinct consumers and authority. Agentic research selects chunks for synthesis, then ordered budget trimming can omit selected sources while retaining an oversized first chunk. See RT-RTE-wiki-admission, MEM-OBJ-wiki-proposals and MEM-OBJ-wiki-versions.
Reuse, warrant and improvement remain separate
Atomic implements cited knowledge products as advertised, but its inspected citation checks establish eligible identities and known reference associations, not that a source supports each assertion. Unknown markers can lose citation rows without rejecting prose. Generated wiki/finding chat readers remove bracketed numeric expressions and cite the product under a new run identity. Genuine evidence-marker removal can preserve a proposition while hiding its original reference; the same deletion can remove an asserted numeric value. Arbitrary delivered prose therefore has no established semantic-preservation guarantee. MCP product delivery is a separate, richer citation channel. See EPI-CLM-cited-products, EPI-RTE-citation-resolution, EPI-RTE-wiki-citations and EPI-RTE-generated-delivery.
The strongest supported trace-fed contribution is the report chain: recorded research messages and tool results feed automatic production of a durable finding available to later readers. This establishes a wired trace-to-product update, including session and tool-trace inputs, independently of novelty, weight updates or benefit. Profile verification accepts that local classification against RT-RTE-report-finding and MEM-OBJ-report-findings. It does not demonstrate improved capacity for future action. Saved-answer derivation is also wired, but its per-save admission control remains unresolved.
These contributions do not establish a theory-building loop. The runtime records do not supply condition-by-condition findings that would warrant declaring theory-builder conditions 1–4 met. Localized prose alone does not identify a formulated theory; delivery does not establish content-dependent consumption; structural rejection does not establish stated content-directed criticism; repeated generation does not establish iteration retaining that criticism. The epistemic account leaves generation semantics and evidence-consuming acceptance unresolved. Likewise, criticism of a consumed theory improving future capacity, reflection, autonomous theory-building and self-improvement are independently unestablished—not inferred absent from partial inspection. Scheduled computation and automatic admission are supported, but do not supply those stronger properties. See epistemic bounded conclusion and RT-RTE-settings-change.
For continuity and corpus-grounded generation, the inspected architecture connects retention, selection and later model consumption. For independently warranted findings or strict content isolation, its present gates support narrower conclusions: citation identity is not factual support, and the report's direct ID reader is not constrained by its search ContextFilter. For improvement assessment, the inspected longitudinal evaluator measures fixture-relative session retrieval through fixed arithmetic, not answer truth or a production update policy. A source-linked execution showing later dependence on a retained change would establish activation; an objective-based intervention and comparison would be needed for improvement attribution. Evidence-consuming acceptance traces would alter the warrant assessment. Applying scope checks to every relevant ID reader, enforcing review across publication alternatives, or retaining full evidence lineage through product delivery would change the corresponding source-level assessments. See RT-RTE-report-finding, workflow account and EPI-RTE-retrieval-scoring.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No registered execution or causal comparison | SRC-1; RT-RTE-chat-turn, RT-RTE-report-finding, EPI-RTE-retrieval-scoring | Frozen implementation and doctrine; no target execution | Successful activation, reliability, retrieval/answer quality and improved future capacity or causal benefit | Provenanced execution artifacts and objective-based comparisons, with independent variation for component attribution |
| Included clients, backends and strategies remain incompletely inspected | SRC-1; RT-RTE-capture-write, RT-RTE-state-maintenance, MEM-OBJ-atoms, MEM-OBJ-wiki-articles | Selected core, SQLite, server/cloud and client paths; alternative admissions, batch/re-embedding, wiki strategies and control payloads unresolved | Complete mechanism/substrate inventories, backend parity, universal consent, isolation, freshness or durability | Route-specific inspection of remaining clients, adapters, strategies and control consumers, including alternate admission and recovery paths |
| Withdrawal and restoration are only partially traced | SRC-1; MEM-OBJ-atoms, MEM-OBJ-wiki-versions | Core/SQLite deletion and archive readers; downstream cleanup, backups and account removal unresolved | Secure erasure, universal withdrawal of copied evidence or automatic rollback | Full cleanup/restore dependency trace across chunks, vectors, citations, archives, backups and adapters, with relevant execution evidence |
| Mixed representational-form units have coarse consumer granularity | SRC-1; MEM-OBJ-chat-history, MEM-OBJ-wiki-proposals; profile verification limit | Conversation text/tool metadata and candidate prose/base metadata; forms are supported but paths differ | Part-level or consumption-path comparisons assuming all fields reach one consumer or receive one interpretation | Split classification units by actual retained part and consumer: provider replay versus storage readers, and review prose versus freshness metadata |
| Generated-product transformations can lose lineage and assertion content | SRC-1; EPI-RTE-generated-delivery, MEM-RTE-answer-save | Chat wiki/finding numeric deletion and frontend answer conversion; separate MCP channel | Faithful arbitrary-prose delivery or transitive source support from a product citation | Candidate-specific input/output and reference mappings, or semantics-aware transformations preserving assertion content and original evidence links |
| Semantic acceptance and theory/improvement mappings remain unresolved | SRC-1; EPI-OBJ-generated-prose, RT-RTE-wiki-admission, RT-RTE-report-finding | Inspected candidate generation, structural admission and later readers; no individuated premises/output or theory-criticism episode | Accepted ampliative knowledge, condition-by-condition theory-building, reflective/autonomous theory-building or evidence-responsive self-improvement | Formulated theory and premises, stated criticism with blame and disposition, retained consequence and later consumer; separately, self-representation/causal-path evidence and objective-based improvement evidence |
| External providers and host operation are not frozen | SRC-1; RT-CMP-agentic-model, RT-CMP-tagging-model, RT-CMP-embedding-model | Configurable inference clients, not immutable weights, service internals or external MCP loops | Provider parameter-learning/version guarantees and end-to-end external-agent behavior | Pinned provider/service evidence and bounded host traces under an explicitly expanded evidence boundary |
The accepted report verification declares no additional limits, and the reconciliation leaves no unresolved conflict. The profile's coarse-unit limit is carried above. Source-read log: none; this synthesis uses the supplied members and adds no source evidence.