Cortex synthesis

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

Bounded synthesis

A knowledge service with caller-owned task progression

Cortex captures typed knowledge objects, maintains relationships and metadata, and returns search results, compact context, documents, dossiers and period narratives. It is not the enclosing agent's planner. This account covers the shipped system in https://github.com/abbacusgroup/cortex at commit a51c2209eb6d26812c19c97aaade3cb64ccfc23c, with applicability cutoff 2026-10-17. The whole-system boundary includes CLI, MCP, REST, dashboard, importers, storage, processing, maintenance, prompts and evaluators. Evidence is inspected implementation and attributed doctrine/design from SRC-1, not observed execution. No target tests, benchmarks, provider calls or services were run, and no execution trace or causal comparison was supplied. External hosts, provider internals and dependency implementations are not frozen evidence. See boundary and evidence and the Source register.

An ordinary CLI capture calls the shipped MCP client and server; the server opens shared storage, processing, retrieval and access-learning services. Fixed Python code owns internal stage transitions. Optional model calls return classifications, relationships or prose; they are not delegated agents with independent tool loops. The external caller chooses whether to capture, retrieve, inspect, give feedback or act. IDs name knowledge objects, not delegated tasks. See RT-RTE-capture-pipeline and the workflow account.

Admission comes before optional interpretation

The representative connected workflow is one capture request with processing enabled: caller input → object creation → normalization → linking → enrichment → reasoning → returned ID and stage results. This is source-supported control flow, not a demonstrated successful run. The object is already retained when classification or relationship processing starts. A supplied nonempty summary skips internal classification; disabling processing skips all four stages. Explicit caller entities override extracted entities, including when the supplied list is empty. Consequently neither internal model review nor entity extraction is a universal admission gate. See RT-RTE-capture-pipeline and the workflow account.

Normalization can propose summary, type, properties and entities from title and at most 8000 body characters. Embedding generation separately uses title and at most 2000 body characters. Relationship discovery sees the new object's bounded content and at most twenty candidate titles/types/IDs, not all supporting documents. Parsing, allowed values, confidence thresholds and endpoint checks constrain proposals. They do not test whether a summary is faithful or a causal relationship is true. In particular, the link checks do not enforce that an endpoint belongs to the prompt's candidate list or is the newly captured object. Caller-provided confidence is not a calibrated correctness probability. See EPI-RTE-metadata-generation, EPI-RTE-structure-check, EPI-RTE-metadata-admission and the workflow account.

Direct CLI and import paths matter. They can open storage outside MCP; importers can create objects without the pipeline, and direct CLI processing omits the embedding provider supplied by MCP construction. Import hash/title checks can skip duplicates, but a hash marker can precede successful object creation. Thus equivalent indexing, universal server-side controls and automatic source-file synchronization do not follow from ordinary capture. See RT-RTE-direct-import.

Retained parts have different consequences

SQLite holds document prose, summaries, metadata, vector bytes, access counters, query logs, weight configuration and historical document rows. Oxigraph holds identities, entities, mentions and relationship assertions. Full reads treat SQLite content as authoritative and add graph relationships. Natural-language bodies inform humans or models; symbolic fields and predicates guide fixed filters and queries; dense vectors affect similarity. Stored vectors are not provider weights. These parts have different lineage, mutation and recovery coverage. See RT-OBJ-store, MEM-OBJ-documents, MEM-OBJ-relationships and MEM-OBJ-embeddings.

Four shipped graph rules derive symmetry, inverse relations and transitive supersession under explicit formal assumptions. For example, encoded A-supersedes-B and B-supersedes-C can yield A-supersedes-C. This is conditional entailment, not independent evidence that the premises or their encoding are true. Derived edges persist and can become later premises, but additive inference does not establish dependency-based retraction after premise changes. Model-generated metadata and period prose remain semantically indeterminate without source/output pairs; retention and operational use do not supply evidence-consuming truth acceptance. See RT-OBJ-inference-rules, RT-BAP-rule-execution, EPI-RTE-rule-derivation, EPI-RTE-rule-retention and EPI-RTE-period-synthesis.

Stage failure normally permits later stages to continue while retaining the original object. Some errors are suppressed locally or converted to fallback results, so even complete does not certify successful writes, semantic quality or rule closure. The returned top-level capture type is initialized before normalization; it is not a fresh read of the final retained classification. Explicit reprocessing does not restore original entity arguments and can reuse a stale or fallback summary. See RT-CLM-stage-resilience, RT-CLM-store-consistency and the workflow account.

Requested reuse controls the response, not the host's context window

Hybrid search combines keyword, optional semantic, graph-connectivity and created-at recency signals. Its default weights are 0.4/0.3/0.2/0.1. Semantic search inspects at most 10000 stored vectors. Context defaults to ten results and projects summaries or title fallbacks plus identifying metadata; a separate ID read returns fuller content and relationships. This keeps document bodies out of a requested briefing, not out of an already accumulated host conversation. Summary length is not hard-capped by the presenter, so result counts do not establish a strict response-token budget. The host still decides whether the returned material enters a model invocation. See RT-RTE-retrieve-context, MEM-OBJ-summary and RT-BAP-knowledge-delivery.

Other presentations use different selectors. Dossier first exact-matches entity names, then falls back to lexical search. Period synthesis filters a bounded object listing and supplies at most twenty summaries/titles to an optional model; its response is not automatically retained. Alerts and advanced reasoning also require a request. Structural dependency checks can support bounded warnings, but repeated fixes do not establish one systemic cause, and absent decision objects do not establish absent architecture documentation. Semantic LLM contradiction comparison is declared in a docstring but not implemented by the inspected contradiction method. Findings do not automatically correct, withdraw or exclude the underlying knowledge. See RT-RTE-retrieve-context, MEM-RTE-issue-delivery, EPI-RTE-period-synthesis, EPI-CLM-semantic-checks and the epistemic ledger.

Usage changes tier metadata; improvement remains a separate question

The strongest supported trace-fed contribution is durable access-derived adaptation: successful full reads and positive feedback increment retained counters, and ten accesses can promote an object to reflex. Later output displays tier, and enrichment preserves that tier. This is a wired activity-to-metadata-to-later-consumer chain; it does not require a new theory or a parameter update. Negative feedback does not reverse the count or withdraw the object. See MEM-OBJ-access-count, MEM-OBJ-tier and MEM-RTE-access-promotion.

The unresolved question is whether this improves future action. Tier and access count are not hybrid-score inputs. Persisted ranking-weight APIs exist, but bounded source searches found no shipped bridge loading those values into retrieval or using feedback to tune them. Query logs have diagnostic consumers, not an established automatic policy-update consumer. Thus automatic ranking adaptation is absent within that searched shipped boundary, while metadata adaptation is wired and its benefit unestablished. Generic captured documents may contain traces, but their unknown original content prevents a complete trace-learning inventory. See MEM-ABS-adaptation-wiring, MEM-OBJ-retrieval-weights, MEM-OBJ-query-log and MEM-CLM-learning-scope.

The independent theory-builder findings concern the formal rules: condition 1, localized stated content, is wired; condition 2, decisions depending on that content, is wired; condition 3, content-directed criticism with revision or changed reliance, remains uninspected; condition 4, retained criticism shaping another round, remains uninspected. Inspectable rule names and bodies do not themselves establish rule criticism. Criticism of a consumed theory improving future capacity is unestablished. Reflection is also unestablished: knowledge-state diagnostics do not establish a two-way representation of Cortex's own method. Internal stage execution is automatic, but autonomous theory building remains uninspected; this is not a verdict on an excluded host's autonomy. See RT-RTE-maintain-knowledge.

For the access pathway, the boundary is Cortex's service and retained metadata across invocations, with the declared objective of better future retrieval. The counter-to-tier change reaches live display and enrichment consumers, but frequency is not independent evidence of usefulness, and no subsequent execution trace closes the causal dependence on that change. Exercised self-improvement is therefore unestablished; even a dispositional improvement-directed claim remains limited by the missing objective-bearing evaluation. This is separate from the supported trace-fed write mechanism, from reflection and from autonomous theory building. See MEM-RTE-access-promotion, MEM-CLM-learning-scope and RT-RTE-maintain-knowledge.

Maintenance and the assessment boundary

For a caller seeking compact reuse of previously captured material, the implementation supports a connected retention-and-retrieval service with optional enrichment and conditional graph inference. For a scenario requiring independently validated claims, universal refresh, contradiction-driven withdrawal or demonstrated retrieval improvement, these records do not establish the needed guarantees. This is a scenario-relative distinction, not a product ranking.

Maintenance remains caller-selected. Direct backing-store writes bypass some temporal snapshots; document edits need not refresh summaries or vectors; current deletion is not demonstrated historical erasure. Doctor checks/rebuilds concern index consistency and graph closure, not source truth, and rule-query failures can appear as zero pending consequences. Backup restoration attempts whole-store replacement rather than semantic rollback. See MEM-OBJ-document-versions, MEM-OBJ-summary, MEM-OBJ-embeddings, EPI-RTE-health-check, EPI-RTE-health-repair and RT-RTE-maintain-knowledge.

Control guarantees are likewise path-specific. Non-loopback HTTP MCP omits admin tools but retains public mutations; REST authentication does not gate the independent MCP listener. The safety warning is not a delete-approval token. Direct local writers remain alternatives. Model adapters request inference, do not pin exact model artifacts and do not inspect provider parameter changes. Package releases and service installations have separate operator/developer admission paths, not a knowledge-driven software-improvement loop. See RT-RTE-transport-control, RT-RTE-install-release and runtime components.

Evidence that would change this assessment includes paired source/model outputs with independent fidelity checks; linked capture-to-retrieval-to-host-action traces; and controlled retrieval comparisons that vary the proposed adaptive mechanism independently. A shipped feedback-to-weight-to-ranking consumer could change the wiring finding without by itself proving benefit. Retained criticism naming a challenged rule, changed reliance or successor rule, and a later criticism-informed round could change the separate theory-builder findings. Coordinated summary/vector refresh, premise-dependent retraction and enforced approval across alternate writers would change the corresponding architectural guarantees, if established by revised records. No new source evidence was added for this synthesis.

Limitations

limitation	affected SRC IDs and record citations	inspected boundary	conclusion prevented	evidence that would resolve it
Imported relationship mapping is coarsely assessed. Exact, normalized or fuzzy wiki-title matching can write supports; the ledger does not separately assess its warrant.	SRC-1; RT-RTE-direct-import; epistemic coverage; report verification limits	Shipped importer and accepted admission/graph account	Complete comparison of imported relationship semantics and checks; title similarity does not warrant substantive support.	A separate mapping/check account and source-link examples evaluated against the intended support relation.
Diagnostic retained-state inventory is incomplete. An in-process tracemalloc snapshot is reused for a later comparison and cleared on start/stop.	SRC-1; memory coverage; report verification limits	MCP diagnostic handlers, beyond detailed memory-body inspection	Exhaustive inventory of all retained diagnostic state or cross-invocation read-back; no ranking or provider-learning inference follows.	Recorded snapshot lifecycle, consumers and read-back classification covering these handlers.
Archive recovery authority is not separately classified in the partial authority inventory.	SRC-1; RT-RTE-maintain-knowledge; profile verification limits	Operator-selected whole-store archive restoration	Consumer-by-consumer recovery-authority comparison; supported storage/form/requested-read findings remain valid.	Classification of the retained archive at its actual restore consumer and replacement effect.
Source-only evidence does not establish operation or benefit.	SRC-1; boundary; RT-RTE-capture-pipeline; MEM-RTE-access-promotion	Frozen repository, no supplied run or experiment; hosts/providers outside boundary	Successful capture/recovery, actual host injection, improved future capacity and exercised self-improvement.	Linked execution artifacts and appropriately scoped outcome comparisons; independent variation for component attribution.
Provenance and model-output meaning remain unresolved.	SRC-1; MEM-OBJ-documents; MEM-OBJ-summary; EPI-RTE-metadata-generation; EPI-RTE-period-synthesis	Generic caller/imported content and optional bounded model prompts	Authenticated source warrant, faithful compression, accepted novel knowledge and complete trace-origin coverage.	Original input provenance, concrete source/output pairs and evidence-consuming acceptance records for named uses.
Refresh, history and withdrawal guarantees differ by mutation path.	SRC-1; RT-CLM-store-consistency; MEM-OBJ-document-versions; MEM-OBJ-embeddings; EPI-RTE-rule-retention	Unified versus backing-store writes, snapshots, additive inference	Universal synchronized state, current projections, historical accuracy, complete erasure and premise-dependent withdrawal.	Coverage of every relevant mutation and consumer, coordinated refresh/retraction and recovery evidence.
Formal-rule consumption does not establish a criticism-and-revision loop.	SRC-1; RT-RTE-maintain-knowledge; RT-OBJ-inference-rules	Fixed rule execution and requested diagnostics	Theory-builder conditions 3–4, criticism-driven future-capacity improvement, reflective or autonomous theory building.	Content-directed criticism with blame, resulting revision/changed reliance, retained next-round effects, and separate reflection/autonomy evidence.
Evaluation and inspection are not exhaustive.	SRC-1; epistemic coverage; MEM-CLM-learning-scope; report verification	Scoped evaluator and implementation reads; synthetic fixtures, unexecuted tests, partly inspected optional surfaces	System-complete evaluation-quality judgment or system-wide negative about unassessed acceptance routes.	Additional material-route/evaluator records with explicit domains and independently inspectable results.