Cortex synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Cortex is a self-hosted memory service for private Markdown notes, with a dashboard, MCP tools and an embedded model reader. It is not the enclosing conversation agent: an external assistant chooses tools and interprets their results. This account covers the shipped service, storage adapters, callers, maintenance instructions and evaluators at commit 65d449a3275eee51efd6ce0c275572806ee9d553, with cutoff 2026-11-07. Its evidence is SRC-1 implementation and doctrine/design, not supplied executions or experiments. The private notes repository, deployed database, provider accounts, evaluator labels and host scheduler are not supplied. The boundary and source register establish that evidence identity and its exclusions.
Admit content and preserve distinct state
Durable note writes become commits in a private GitHub repository. Optional Postgres stores a rebuildable note mirror, deliberate working-state items and serving-access history; optional KV stores pending proposals and reusable answers. Working state is not recoverable from the notes repository. Note prose supplies evidence or context to later readers, while paths, lifecycle fields, scores and edge weights have fixed-rule consumers. Numerical ranking metadata is symbolic state, not learned model parameters. See MEM-OBJ-notes, MEM-OBJ-mirror, MEM-OBJ-bubble-body, MEM-OBJ-bubble-state and MEM-OBJ-cached-answer.
Admission depends on the route. A trusted credentialed caller can write, append, capture or replace notes and update working state. Code checks paths, sizes, modes and exact edit targets; it does not require evidence that the new assertion is true or per-content human approval. Local import and reorganization scripts have separate batch gates, and local file changes require publication before production consumption. Imported text explicitly lacks verification history. Guest proposals instead wait for a trusted reviewer, which need not universally be a human. Authorization and persistence therefore do not by themselves establish evidence-consuming acceptance of a claim. See RT-RTE-trusted-write, MEM-RTE-local-curation and RT-RTE-proposal-admission.
Recovery is also route-specific. Proposal acceptance publishes the target and an operation receipt in one Git commit through a non-forced ref update. Exact-ID retries can recover the committed outcome even after queue cleanup, assuming intact receipts and branch history. Ordinary writes have no equivalent durable replay key: reply loss can follow a successful commit, and index regeneration can fail after the note is saved. Filing a working item is not an atomic transfer into a durable note. These distinctions matter when resuming interrupted work; a missing response is not proof of no effect. See RT-RTE-proposal-admission, RT-RTE-trusted-write and MEM-OBJ-bubble-state.
Select context before the host or reader consumes it
Cortex offers both direct retained-text delivery and a returning answer computation. Boot context, project handoff, corpus selection and exact reads serve an external conversation without inheriting its whole window. Boot preferentially includes a profile head, router descriptions, recent material and usable working state. It omits cold router entries, aggregates logs by month, clips bodies and applies section budgets. A usable bubble suppresses expanded unscoped daily logs; scoped boot still selects matching project log sections. Handoff gathers project material and ranked neighbors under its own byte budget. Omitted bodies remain reachable by fuller reads; pointers do not automatically reload them. This is service-side working-context control, not Cortex-owned host compaction or a launched subagent frame. See RT-RTE-context-delivery and MEM-OBJ-router.
For Ask, code loads a corpus snapshot, resolves policy/model settings and checks an optional answer cache. On a miss, lexical selection normally chooses up to 15 notes under a 400,000-byte body budget, with additional log and page-part caps. Full mode is still budgeted, not an unconditional complete read. The provider receives an answer contract, selected note blocks and the question; it proposes prose and citation, while server code owns packing, checking and rendering. The reader has no tool-selection or recursive execution loop. Its allowlisted provider model names do not pin immutable weights, and inspected serving code does not update those weights. See RT-RTE-reader-answer, RT-OBJ-answer-contract and RT-CMP-reader.
The guest door registers only Ask and Propose. Its scope filter precedes selection and citation checking on cache misses; readable policy is required even on hits, and policy identity separates cached replies. The unrestricted empty-scope library branch is not a demonstrated guest escape: the shipped policy writer rejects empty scope and its parser supplies defaults. Misses spend an atomic KV-backed daily model-call counter; hits intentionally bypass inference and spending. Citation fields are suppressed by default, but that is not a formal non-disclosure constraint on answer prose. The dashboard Ask is a separate trusted consumer with a process-local ceiling, not a distributed guest budget. See RT-RTE-guest-mediation and the runtime account.
Check the evidence actually claimed
A positive Ask verdict checks normalized quote occurrence within an issued-tag source block at the loaded commit. It does not check whether the answer follows from that quote or whether the note is true. Correction/retraction markers and ambiguous matches change labels, not this limited license. UNVERIFIED prose can remain visible. NOT IN BRAIN requires the relevant abstention/coverage branch, but coverage can mean that no unread note shares a question token; it is not semantic exhaustion. Timeouts and notes that do not fit are not treated as proof of absence. Direct context/read routes do not receive Ask's verdict. See RT-CLM-quote-verification, EPI-RTE-quote-check and EPI-RTE-answer-stamp.
When enabled, the cache admits non-UNVERIFIED rendered replies, including negative or superseded verdicts. Matching question, corpus SHA, resolved model, selection and trust policy can return a stored reply without a fresh model call or renewed verification. Replay preserves the original verdict's limits rather than upgrading its warrant. Likewise, watch diagnostics establish bounded structural predicates, not general semantic criticism: reciprocal correction-marker references are review candidates, not proof that pages contradict one another or have temporally re-corrected each other. See EPI-RTE-answer-cache, MEM-OBJ-cached-answer and EPI-CLM-correction-chain.
Adapt selection and maintain content
The strongest implemented adaptive contribution is an automatic trace-to-selector chain: serving tools append access events; SQL derives decayed temperatures and coaccess edges; later router and handoff selectors consume them. This can alter later context without changing note prose or model weights. Access is evidence of delivery, not cognitive use or successful action. Observation writes can be dropped, graph rebuilds can refuse publication, and handoff edge reads do not guarantee fresh graph state. Maintenance exclusion uses a caller-declared mode; it does not authenticate the origin of every maintenance read. This supports wired trace-fed behavior shaping, not observed improvement in future capacity. See MEM-RTE-usage-selection, MEM-OBJ-note-access, MEM-OBJ-note-scores and MEM-OBJ-note-edges.
The representative multi-step agent workflow is the shipped groundskeeper template. An operator customizes and schedules it in an external Claude Code host. The host first runs a connector/tool-roster healthcheck and requests boot context; either failure calls for no writes and an external report. On the healthy branch it reads recent logs, files selected durable facts, preserves logs with absorption markers, checks at most four drift pages against live evidence, reports pending proposals, and captures a short digest when something merits reporting. Quiet healthy runs deliberately write nothing. The host owns semantic judgment, sequencing and recovery across calls; Cortex supplies structural edit rejection and individual persistence receipts. The template reserves proposal accept/reject for a human, but that restriction is not enforced on all trusted writers. No supplied execution closes the workflow's handoffs. See RT-RTE-groundskeeper-maintenance, RT-BAP-maintenance-instructions and the workflow account.
These maintenance instructions request live-evidence checking, visible corrections and uncertainty marking for later cross-device use. They are the closest described route to evidence-consuming fact acceptance, but do not demonstrate that checking, acceptance or integration occurred. The activity-log-to-page/digest path remains claimed rather than an exercised host loop. Actual input/output pairs would also be needed to distinguish faithful reshaping from new conjecture in reader answers or maintenance outputs. See EPI-CLM-drift-check and the epistemic bounded conclusion.
Assess the supported uses, not a global grade
For a client requesting bounded resumption context, the implemented contribution is selective delivery with fuller-read escape routes, not complete automatic recall. For a client requiring grounded answers, the implemented contribution is source-occurrence evidence and explicit verdicts, not factual certification. For unattended maintenance, the contribution is a conservative host procedure backed by durable tools, not a demonstrated reliable nightly service. For usage-sensitive retrieval, the contribution is a wired retained adaptation path; its usefulness remains unmeasured in the supplied evidence.
The evaluator families address different questions. Reader evaluation compares answers with external labels and model judges; retrieval evaluation checks expected paths and phrases, with an optional local recall-floor failure; prediction replay compares rankings against hidden access windows. Prediction replay uses current structural information rather than a fully historical corpus, and does not install a winner. No supplied labels, results or intervention establish answer quality, ranking benefit or a component effect. See RT-RTE-reader-evaluation, EPI-RTE-retrieval-evaluation, EPI-RTE-prediction-evaluation and MEM-CLM-evaluation-gate.
Learning, reflection, theory building and self-improvement require separate conclusions. Localized behavior-directing method text is available, but a complete formulated-theory route is not established. The reports do not establish content-directed criticism with blame placement and resulting revised reliance, nor retention of that criticism shaping another round; they therefore do not establish all four theory-builder conditions. Nor do they establish that criticism of consumed theory improved future capacity. Reflection's two causal directions and role-complete autonomous theory building remain unestablished, independently of automatic graph updates. The service-owned reader is a returning computation, while the multi-step maintenance agent belongs to an external host. These are bounded conclusions, not absences across every shipped development path. See the epistemic bounded conclusion and report-verification limits.
For self-improvement, the usage chain supplies a plausible organizational update and later ranking authority, but no subsequent operation trace establishes causal dependence on an objective-directed update or successful improvement. Groundskeeper writes and operator-confirmed deployment changes do not close those links either. Console production jobs require signed preparation/confirmation and identity-checked reconciliation; local updates are an interactive alternative. Provider acceptance is not successful execution, and neither gate covers every deployment route. Delegated CI repair and PR review remain only partially characterized. See MEM-RTE-usage-selection, RT-RTE-production-jobs, RT-RTE-local-update and report-verification limits.
The assessment would change with retained traces connecting admitted content or derived metadata to later consumption and action, label-bound evaluation results, and comparisons isolating the claimed adaptation's effect. A semantically stronger answer guarantee would require an additional evidence-consuming entailment or truth check, not merely a stronger interpretation of VERIFIED. Deterministic nightly replay claims would require aligned completion markers and a cross-call recovery protocol. No frozen-source reads were needed for this synthesis; all findings use the accepted members.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No deployment, private corpus, run or causal comparison supplied | SRC-1; RT-RTE-reader-answer, MEM-RTE-usage-selection, RT-RTE-reader-evaluation | Frozen service, adapters and evaluator source; private state and results unavailable | Observed activation, reliable recall/answers, successful maintenance, improved capacity and exercised self-improvement | Provenance-linked execution artifacts and later-consumer traces; scoped evaluation results and controlled comparisons |
| Development-plane coverage is partial; carried report-verification limit | SRC-1; RT-RTE-production-jobs, RT-RTE-local-update, epistemic conclusion, verification limit | Service/deployment routes characterized; delegated CI repair and PR review in .github/workflows/ci-autofix.yml and .github/workflows/claude-review.yml only partially covered; external reusable implementation not frozen |
Complete comparison of code-repair/admission automation; every application-change proposer being human; successful external repair or exercised improvement | A new bounded analysis with the external implementation frozen and route/operation evidence |
| Groundskeeper marker spellings differ; carried report-verification limit | SRC-1; RT-RTE-groundskeeper-maintenance, workflow account, verification limit | Complete maintenance template: skip closed undated token, append dated token; individual write implementation | Deterministic replay suppression or exactly-once nightly filing; host interpretation may reconcile intent but is unobserved | Aligned marker predicate/output plus demonstrated interruption/retry behavior; cross-call completion protocol for an exactly-once claim |
| Health, semantic verification and completion remain host-owned | SRC-1; RT-RTE-groundskeeper-maintenance, EPI-CLM-drift-check, workflow account | Template, roster healthcheck and individual writes; customized host and evidence grants unavailable | Guaranteed scheduling, enforced health prerequisite on every writer, completed human review or accepted fact corrections | Installed task/configuration, evidence-bearing check/disposition records and complete handoff traces |
| Initial content, maintenance-log origin and administrative retained payloads remain incompletely classified | SRC-1; MEM-OBJ-notes, MEM-RTE-local-curation, RT-RTE-groundskeeper-maintenance, RT-RTE-capability-configuration, profile verification | Known subsequent writes and consumers; seed/private provenance and detailed operational retention unresolved | Exhaustive memory inventory, universal write-agency or trace-source classification, or system-wide absence of learning | Recorded initialization/admission lineage and complete payload-to-later-consumer accounts |
| Freshness and recovery differ by state plane | SRC-1; MEM-OBJ-mirror, MEM-OBJ-note-edges, MEM-OBJ-bubble-state, RT-RTE-trusted-write | Cached corpus fallback, separate graph/bubble reads, ordinary versus receipt-backed writes | Universal current snapshot, deployment-level recovery guarantee or no-effect inference after reply loss | Deployed schema/store contracts, state-identity traces and route-specific recovery evidence |
| Access history is a delivery proxy; mode and age predicates are narrower than prose labels | SRC-1; MEM-RTE-usage-selection, MEM-OBJ-note-scores, MEM-RTE-retirement-nomination | Caller-declared maintenance mode, bounded score history and nomination-only CLI | Successful-use learning, exclusion of every actual maintenance read, lifetime nonuse or executed deletion | Authenticated work-origin/use outcomes, appropriate lifetime history and a traced retirement decision/consumer |
| Quote and diagnostic predicates have limited semantic license | SRC-1; EPI-RTE-quote-check, EPI-RTE-answer-stamp, EPI-RTE-answer-cache, EPI-CLM-correction-chain | Occurrence, lexical coverage, cache reuse and reciprocal marker checks | Answer truth/entailment, semantic exhaustion, renewed warrant on replay or diagnosed logical contradiction | Separate semantic evidence-consuming checks within a named domain; actual source/answer or proposition comparisons |
| Provider and host internals are outside frozen implementation | SRC-1; RT-CMP-reader, RT-RTE-context-delivery, RT-RTE-guest-mediation | Provider adapters and service selection/guest controls, not remote weights or external conversation management | Immutable provider behavior, host compliance/compaction guarantees or deployment-wide confidentiality | Pinned provider artifacts/contracts and host/deployment evidence matching the claimed guarantee |