Stash reconciliation

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

Reconciliation

Evidence and responsibilities

The four reports share the run identity and frozen revision in boundary. Their evidence is implementation and doctrine/design from SRC-1 in the Source register, not observed operation or causal experiments. Runtime describes a memory service with a maintenance controller, not an enclosing agent's task loop. Memory and epistemic add detailed inspection within that same source boundary. Workflow traces one instructed session-memory sequence and declares no new records. This reconciliation uses only the supplied members; no target source was read.

Duplicate supersessions

These connections identify the same source-native referents, not interchangeable findings. Both declarations remain available. No supersession changes a status, supplies missing evidence or turns stored content into accepted truth.

Amendment: EPI-OBJ-episode is superseded by MEM-OBJ-episode. Both identify caller assertion text stored as Episode content in PostgreSQL and consumed by recall and synthesis, supported by SRC-1 internal/brain/episode.go. Memory additionally describes event metadata and repeat-failure production. The affected epistemic acquisition and retention findings concern this same text, not an authenticated execution trace. The separately declared numerical vector remains distinct.

Amendment: EPI-OBJ-fact is superseded by MEM-OBJ-fact. Both identify Fact content, entity/property/value and confidence/validity/source metadata, with synthesis, failure-pattern and hypothesis producers. SRC-1 internal/models/models.go and the admission paths cited in the records identify the same stored belief. The affected object inventory and ledger references retain producer-specific warrant. MEM-OBJ-memory-vectors separately owns the numerical selection representation; this amendment does not merge it with text or provider weights.

Amendment: EPI-OBJ-relationship is superseded by MEM-OBJ-relationship. Both identify Relationship triples with source_fact_id and confidence in relationships, consumed by queries and pattern extraction. Their SRC-1 model declaration and extraction/storage evidence concern the same persisted edge. The affected proposition inventory acquires neither causal nor explanatory warrant from this identity.

Amendment: EPI-OBJ-pattern is superseded by MEM-OBJ-pattern. Both identify Pattern content, source arrays, coherence and confidence in patterns, supported by their SRC-1 model and consolidation evidence. The affected inventory concerns abstraction rows, not failure-pattern facts. QueryPatterns capability and actual shipped consumption remain separate questions.

Amendment: EPI-OBJ-causal-link is superseded by MEM-OBJ-causal-link. Both identify CausalLink directed fact-ID edges with method and confidence, returned through list/trace. Their SRC-1 model and internal/brain/causal.go evidence identify the same structure. The affected causal-object findings remain assertions rather than causal-experiment results.

Amendment: EPI-OBJ-hypothesis is superseded by MEM-OBJ-hypothesis. Both identify Hypothesis content, plan, rejection reason, status, confidence, source IDs and confirmed-fact link in PostgreSQL. Their SRC-1 model/lifecycle/formatter evidence concerns the same candidate. The affected ledger references preserve the automatic evaluator's omission of the plan and the difference between disposition and empirical testing.

Amendment: EPI-OBJ-contradiction is superseded by MEM-OBJ-contradiction. Both identify Contradiction old/new values, fact IDs, confidence, resolved flag and resolution text. The SRC-1 model declaration and contradiction writes identify the same row. The affected object inventory retains the omission of automatic model explanations from persistence.

Amendment: EPI-OBJ-goal is superseded by MEM-OBJ-goal. Both identify Goal intended outcome, notes and symbolic status/parent/priority metadata in PostgreSQL, using SRC-1 model and goal-maintenance evidence. The affected inventory still distinguishes intention from an assertion of achievement or progress.

Amendment: EPI-OBJ-failure is superseded by MEM-OBJ-failure. Both identify Failure content, reason and lesson with namespace/goal metadata, using SRC-1 model and failure CRUD evidence. The affected review inputs retain unknown source warrant; neither declaration authenticates a reported execution outcome.

Amendment: EPI-RTE-pattern-extraction is superseded by MEM-RTE-pattern-extraction. Both trace checkpoint-selected facts and relationships through the model, reference/count checks and pattern insertion, with QueryPatterns as the read capability. Both retain SRC-1 internal/reasoner/openai.go's minimum-reference predicate and internal/brain/consolidate.go's source-array fallback. Epistemic additionally describes repeated IDs; memory separately labels read-back afforded. The affected ledger functions remain transformation, checking, disposition and retention, not demonstrated use or accepted ampliative knowledge.

Amendment: EPI-RTE-failure-review is superseded by MEM-RTE-failure-derivation. Both include caller failure admission and later model review against recent episodes, producing unembedded repeat episodes and embedded pattern facts. Their retained SRC-1 internal/brain/consolidate_failure.go passages identify those same outputs and later consumers. The affected review functions retain distinct input checks, output admissions and checkpoint consequences. Neither route establishes authenticated traces, common fact provenance or learning benefit.

Amendment: EPI-CLM-fact-warrant is superseded by MEM-CLM-verified-facts. Both declare the protocol's verified/consolidated-truth assertion for facts and quote the same query_facts wording from SRC-1 cmd/cli/mcp_prompts.tmpl. Epistemic also retains the distilled-truth description. The affected claim comparison concerns this warrant assertion, not an independent verification result. RT-CLM-session-protocol instead concerns requested call order and session behavior.

Containers and overlaps that are not identities

RT-RTE-consolidation-run remains the eight-stage invocation. EPI-RTE-fact-synthesis isolates its stage-one admission witness. Their common source path and witness do not make the part identical to the container. MEM-OBJ-consolidation-progress is its persisted selector state, not another invocation. No split replacement is needed for this valid container.

MEM-RTE-edge-extraction includes relationship extraction, automatic causal extraction, direct detection and caller assertion. EPI-RTE-relationship-extraction covers relationship production; EPI-RTE-causal-links covers causal production/assertion and list/trace consumption. Neither epistemic referent equals the combined memory referent, and their union is not a replacement required by an incorrect combined finding. Different checks and consumers remain explicit. The memory correction removes containment of the whole edge record in consolidation: its direct detector/assertion paths extend beyond maintenance. Only the automatic branch is contained. No current dependent finding requires the removed whole-route containment. The shared LastFactID finding qualifies automatic causal coverage without denying direct admissions or runtime's ordered stages.

MEM-RTE-hypothesis-maintenance combines caller lifecycle control and automatic evidence scanning. EPI-RTE-hypothesis-assessment isolates the scan; EPI-RTE-hypothesis-control isolates caller operations. These are differently scoped overlaps, not duplicate identities. Automatic assessment consumes retained fact text; direct confirmation takes an ID and promotes unchanged content. Epistemic's corrected interface account distinguishes CLI's direct testing transition from MCP create/confirm/reject: MCP confirmation needs testing status obtained elsewhere. Memory's reference to relevant API/CLI/tool surfaces does not assert that every transition is exposed by each interface. Runtime's mention of refinement establishes registration, not successful editing; all detailed accounts preserve the UTC-method-value defect. No dependent finding requires an MCP-only complete direct lifecycle or successful refinement.

MEM-RTE-goal-maintenance includes initial admission and editing as well as model annotation and caller completion. EPI-RTE-goal-assessment covers annotation and completion/status consequences while leaving creation adapters incompletely inspected. Their scopes overlap without identity. Both distinguish model suggestions from caller completion. Epistemic's sibling-status denominator detail qualifies ancestor aggregation rather than contradicting memory's completion account. The entire caller/automatic referent is not contained in consolidation.

MEM-RTE-fact-withdrawal combines manual withdrawal, decay and contradiction handling. EPI-RTE-memory-withdrawal and EPI-RTE-contradiction-review separate decay/deletion/selection from value comparison and resolution. None has the same full referent as another. Their valid overlapping accounts need no split supersession. Caller resolution and withdrawal are not wholly contained in maintenance. All retain validity-insensitive recall and maintenance fetch; resolving a conflict does not repair either fact. Memory's searched FK details and epistemic's unaudited cascades do not conflict: neither claims complete derivative erasure.

MEM-RTE-typed-read differs from RT-RTE-recall-read in selectors and progression: typed filters/traversal versus query embedding and facts-first allocation. Bare getters and QueryPatterns remain callable storage capabilities without a found shipped consumer. Epistemic's pattern-route phrase “API consumer” names that callable interface, while its limits withhold actual downstream use. This does not contradict memory's afforded read-back status; no stronger common status is selected.

MEM-OBJ-context is the retained focus payload of RT-RTE-context-handoff, not a route duplicate. Both distinguish expiry metadata from getter enforcement. MEM-OBJ-namespace, MEM-OBJ-settings, MEM-OBJ-embedding-cache and MEM-OBJ-memory-vectors have no runtime object counterparts. In particular, RT-CMP-openai-embedder describes the adapter/provider model boundary, not its accumulated cache or emitted vectors. No identity is inferred between a producer and its output/store.

MEM-BAP-maintenance-input carries retained evidence to maintenance models; RT-BAP-extraction-policy carries static instructions. MEM-BAP-retained-advice carries retrieved content to an external agent; RT-BAP-agent-protocol carries the static server description. Different payloads and authority roles rule out identity in either pair. The protocol object, its authority path and its call-order claim also have distinct referents. EPI-CLM-memory-learning and EPI-CLM-hypothesis-verification have no duplicate memory/runtime claim records: they concern benefit and hypothesis verification/promotion, not fact labels or call order.

Integration dispositions and independent convergence

The memory integration issues qualify runtime's expressly scoped stage-one witness rather than overturn it. Failure-pattern and hypothesis fact admissions bypass its extraction/provenance checks. Shared causal cursors qualify automatic input coverage. Unregistered PurgeFact dispatch and no found QueryPatterns caller limit interface-wiring conclusions. Best-effort associations, broadened pattern source arrays and dropped model reasons prevent a uniform retained-rationale finding. These findings remain beside their individual records; they do not imply failure of every maintenance stage or complete absence of evidence-sensitive evaluation.

The workflow account, RT-RTE-consolidation-run and EPI-RTE-fact-synthesis independently connect structural admission, partial persistence and recovery. Workflow distinguishes local response retry from later batch replay, which requires another caller/ticker invocation. Embedded stage errors need not activate the external agent's retry/disclosure protocol. None promises atomic rollback, exactly-once processing, complete provenance or successful contradiction review for every fact.

RT-RTE-recall-read, its memory annotations and workflow's later-delivery account agree on requested selected tool text, not global joint fact/episode ranking or delivery of the full derivation. No report assigns the external client's message eviction, obedience or answer composition to Stash. These common limits leave implemented storage, transformation and delivery intact while withholding observed activation and benefit.

Remaining cross-report assessment

Unresolved conflict: MEM-RTE-hypothesis-maintenance labels theory-builder condition 3 claimed, based on requested contradiction and retained rejection reasoning; EPI-RTE-hypothesis-assessment labels the same condition uninspected, because no criticism instance establishes challenged content, result and placement of blame, and explicit blame assignment is not required by implementation. Their SRC-1 hypothesis formatter and maintenance passages support statement-dependent assessment and retained rejection reasons; neither supplies an instance meeting all criticism requirements. Both faithfully retain that uncertainty. The discrepancy is therefore in assessment labels for the overlapping automatic route, not evidence that one report found actual formulated criticism and the other denied it. It prevents reporting one agreed condition-3 status. This reconciliation selects neither label, does not upgrade a requested model behavior into demonstrated criticism, and leaves any required correction to the verifier and declaring analyst. It does not retract their independent positive findings for conditions 2 and 4 or establish a system-wide theory-builder negative.

No correction-answer or declined-objection files were supplied. The corrected containment and interface findings visible in the pinned reports have been assessed above, rather than treating their correction descriptions as established judgments. No other cross-report disagreement was identified. Unknown human control, candidate content, provider versions, external activation and improved capacity remain scoped uncertainties, not additional conflicts. No missing part requiring a new analyst declaration was identified from these members.