KHMS memory reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Connected account and evidence boundary

The reports share the identity and evidence limits of boundary. Its Source register supplies implementation, doctrine/design and reported operation, not an observed-run or causal-experiment layer. The reports connect capture, proposal, review/import and later availability without establishing an exercised successful cycle.

The central progression is RT-RTE-nightly-proposals → RT-RTE-mark-evaluation → RT-RTE-card-admission → RT-RTE-recall-return. Memory identifies retained inputs, intermediate proposals and later consumers. Workflow traces handoffs, checks, recovery and delivery using those runtime records. Epistemic separates the consequential functions within that progression; its ledger does not convert operational import into demonstrated epistemic acceptance.

In particular, the current epistemic ledger separates acquisition of external source packets from generation of one-line findings within RT-RTE-nightly-proposals. The generated findings in MEM-OBJ-inbox have an indeterminate transformation relation, rather than inheriting the acquisition classification of their inputs. This agrees with memory's distinction between raw evidence, model interpretation and software admission of generated lines. Source-span occurrence supports located provenance, not semantic equivalence or an accepted novel claim. No connection from the earlier reconciliation is carried forward as a claim that all nightly content production is acquisition.

Identity and supersession dispositions

No equality requiring a duplicate supersession is established. The overlapping referents compare as follows; shared files or consumers alone do not establish identity.

• EPI-OBJ-mark covers journal/source records and staged packets; MEM-OBJ-trace-material covers raw transcripts, journal MARKs and lossy digests. Their acquisition scopes overlap without establishing equal referents or containment in exactly one supplied object. Neither supersedes the other or the producing route.
• EPI-OBJ-candidate identifies card proposals within the broader generated-line, candidate and reject collection of MEM-OBJ-inbox. The narrower object is not identical to the collection. The container remains valid; no split supersession follows.
• EPI-OBJ-card is a combined card-file object that explicitly distinguishes body and metadata consumers. MEM-OBJ-card-content and MEM-OBJ-card-metadata assess those operative parts separately; the content record also covers archived bodies. These distinctions do not make the combined record's findings wrong, so neither equality nor split supersession is established.
• EPI-OBJ-verdict, MEM-OBJ-mark-ledger and MEM-OBJ-hook-state retain distinct judgment, queue/review and delivery-state referents. Shared file/JSON substrates do not equate their contents or authority.
• MEM-OBJ-embedding-index is retained numerical output, not either producer RT-CMP-qwen or RT-CMP-bge. RT-CMP-agent-model is a separate inference component. No component/output supersession applies.
• MEM-RTE-pending-evidence and MEM-RTE-precheck-return are declared parts of RT-RTE-recall-return, with distinct retained inputs and selectors. They do not replace their valid container or each other.
• EPI-RTE-supersession-link is an integration part of RT-RTE-card-admission. EPI-RTE-validity-check and EPI-RTE-triage-verdict separate evaluation and disposition within RT-RTE-maintenance-admission. These valid containers and parts coexist without supersession.
• EPI-RTE-quote-check spans proposal-stage checking, not either whole proposal route or the separate configurable format checker. EPI-RTE-hook-gates concerns capture/dispatch denial, not merely recall delivery. Neither shares an identical referent with the closest runtime route.
• MEM-RTE-index-bootstrap, MEM-RTE-retired-lookup and MEM-RTE-export-copy have distinct navigation, retired-claim and portable-copy consumers. They do not replace recall or maintenance. MEM-OBJ-views and MEM-OBJ-tombstones identify retained materials, not duplicate routes.
• MEM-CLM-evaluation-scope and EPI-RTE-retrieval-eval concern the same evaluator but declare a scope claim and a comparison route, respectively, not identical records.
• RT-CLM-review-boundary, EPI-CLM-grounding-check, EPI-CLM-synthesis-review and EPI-CLM-measured-injection assert different review, textual-grounding, synthesis and reported-measurement claims. None supersedes another. Workflow declares no new records.

Integration issues and independent convergence

Bootstrap doctrine and generated content. Memory's first integration issue relates MEM-OBJ-views and MEM-RTE-index-bootstrap to RT-BAP-context-delivery. Runtime quotes AGENTS' bootstrap-body inclusion but explicitly calls that channel prescribed. Memory's build_views evidence describes navigation/search instructions without card IDs or bootstrap body. This is doctrine/implementation divergence, not incompatible report findings on the same evidence layer. Neither implemented bootstrap-body delivery nor actual persistent host loading follows from runtime's quotation. The authority path remains valid for recall/hook delivery.

Matched events, history and correction pointers. Memory's second through fourth issues qualify the scope of RT-RTE-recall-return. MEM-OBJ-hook-state distinguishes SessionStart reminders from ranked card search. The parent route does not require every matched event to search. MEM-OBJ-card-metadata restricts history to dead cards still loaded from know/, not archive loading; runtime establishes flagged dead-card history without claiming archive inclusion. Memory's recall annotation bounds and qualifies correction pointers, whereas runtime claims correction-aware selection, not unconditional delivery of every correction. These distinctions prevent universal event-search, archive-retrieval and complete-rationale conclusions without invalidating the recall container.

Recovery alternatives. Memory's fifth issue supplies optional prewrite archive copies in MEM-OBJ-tombstones. RT-RTE-card-admission and RT-RTE-maintenance-admission describe receipt reconciliation and external-version recovery, while denying universal transactional undo. The local-copy alternative does not contradict that denial. Both accounts remain, without making external versions the exclusive possible backup or copies proof of successful restoration.

Different checks have different force. EPI-RTE-quote-check and workflow's nightly annotation independently retain the tolerated quote-check exit. Workflow adds tolerated relation reports, skipped checking after nearest-card preparation failure and header quarantine. This supplements RT-RTE-nightly-proposals; its checked-carry wording cannot establish that every check passed. Runtime's RT-RTE-mark-evaluation already distinguishes optional format repair from downstream checking. Generated-line admission, format repair, quote reporting, publication verification and card import are separate control points. No report's strongest gate supplies force to every stage.

Review protocol and hook enforcement. Runtime, memory and workflow converge on separately invoked reviewed import, not authenticated per-content human approval or universal filesystem confinement. EPI-RTE-hook-gates adds conditional tool-attempt denial; this differs from the model-interpreted context force of RT-BAP-context-delivery. The different consumers explain the different force. Runtime and epistemic agree that Agent/Task dispatch is not wired in the example settings, errors fail open, and report handling does not refuse reports. Neither establishes an unavoidable truth or review gate.

Evaluation and retained semantic limits. MEM-CLM-evaluation-scope and EPI-RTE-retrieval-eval converge on an expected-card oracle, not a final-answer or truth oracle. EPI-RTE-validity-check adds reproduction and aggregate-verdict limits to maintenance, not a replacement maintenance account. EPI-CLM-synthesis-review supports requested criticism and reflection-oriented proposal machinery. Memory's trace-fed reuse and workflow's isolated handoffs establish mechanisms alongside unavailable accepted-claim, compliant-use and benefit evidence. These independently scoped uncertainties are not disagreements.

Memory's remaining integration issues concern the nonduplicate scopes above and unavailable deployment/semantic facts. Epistemic also expressly leaves legacy lexical coverage repair unassessed; that limits evaluator completeness rather than contradicting the supported routes. No analyst correction-answer files or declined objections were supplied, so no correction or decline is inferred from references to a refusal within a member. The present reports establish no unresolved cross-report disagreement on these scoped readings. This reconciliation does not choose stronger statuses or settle the source's doctrine/implementation divergences. No frozen-source read was needed; all connections use the supplied members and their retained evidence.