Cognee reconciliation

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

Reconciliation

The four reports share the boundary and its Source register. They describe inspected implementation and doctrine/design, not observed operation or causal benefit. Added specialist inspection does not convert the boundary into whole-system coverage. No correction answers or declined objections were supplied. This reconciliation uses the supplied members only; no additional source reads were needed.

Duplicate identities

Amendment: EPI-OBJ-learning-document is superseded by MEM-OBJ-session-learning. Both identify the same rendered Session learning document from SRC-1, cognee/modules/session_distillation/distill.py: accepted statement plus optional why_learned prose and session heading, passed to add/cognify for retained graph/vector use. The rendering passages in the memory object and EPI-RTE-lesson-retention establish identity. This does not equate the transient writer decision with its rendered document. Findings in the epistemic inventory and retention/integration ledger refer to the memory object's retained product; their separate epistemic functions and limits remain unchanged. Neither account establishes durable storage success or complete decision provenance.

Amendment: EPI-OBJ-truth-coordinates is superseded by MEM-OBJ-truth-state. Both include persisted DocumentChunk truth_alignment/truth_epoch and vector centroids from SRC-1, cognee/modules/truth_subspace/build.py. Their consumer identity is supported by the same epoch/multiplier passage in EPI-RTE-truth-reranking and MEM-RTE-truth-maintenance. The epistemic inventory and ranking ledger therefore concern the memory object's numerical state. The supersession changes no ranking, recovery or warrant finding. Both distinguish numerical geometry from provider weights and from factual verification.

Containers, parts and distinct identities

RT-RTE-permanent-memory remains a valid ingestion container. MEM-OBJ-graph-content, MEM-OBJ-graph-structure and MEM-OBJ-vector-index identify different retained content, symbolic structure and retrieval representations, not duplicate pipelines. RT-CMP-embedding-config identifies inference configuration, not either the vector index or truth state. RT-CMP-llm-gateway likewise identifies provider dispatch rather than generated content. These identities are not supersessions.

The optional contradiction and temporal parts extend the ingestion account without replacing it. EPI-OBJ-contradiction-list, EPI-RTE-contradiction-check and EPI-RTE-contradiction-annotation separate model judgment from stored warning edges. EPI-OBJ-temporal-edge and EPI-RTE-temporal-supersession identify recency metadata and its deterministic update. The supplied passages establish different checks and consequences: neither conflict annotation nor recency selects empirical truth. Their downstream interpretation remains uninspected, not a cross-report disagreement or system-wide absence.

RT-RTE-session-improve remains the storage/maintenance container. Its parts MEM-RTE-trace-promotion, MEM-RTE-session-distillation, MEM-RTE-ranking-maintenance, MEM-RTE-truth-maintenance and MEM-RTE-summary-readback have distinct transformations and consumers. The cache objects MEM-OBJ-session-qa, MEM-OBJ-session-agent-trace, MEM-OBJ-session-feedback and MEM-OBJ-session-context are retained inputs/results, not alternative declarations of orchestration. No container split requires supersession.

The epistemic distillation records divide the memory route by function, not identity: EPI-RTE-lesson-curation, EPI-RTE-lesson-check, EPI-RTE-lesson-admission, EPI-RTE-lesson-retention, EPI-RTE-lesson-integration and EPI-RTE-distillation-recovery trace stages of MEM-RTE-session-distillation. Their declared improve parent remains valid; the memory route also names independent public distillation entry. EPI-OBJ-proposed-lesson and EPI-OBJ-written-lesson are distinct transient proposal and decision payloads, not the rendered retained lesson. The writer's evidence-consuming judgment and code publication filter coexist with the memory finding that no independent truth oracle is established. No supersession or disagreement follows from the different analytical functions.

MEM-RTE-truth-maintenance builds and commits geometry; EPI-RTE-truth-reranking identifies its later ranking consumer inside recall. These are distinct production and consumption functions, not duplicate routes. EPI-CLM-epoch-recovery concerns coordinate/basis consistency, whereas RT-CLM-improve-coordination concerns process-local exclusion. Their identities remain separate.

WF-RTE-skill-loading is the handler subroute of RT-RTE-agentic-tools, not the entire loop or RT-BAP-skill-guidance's consumer/channel/force/horizon. MEM-OBJ-skill-procedure, MEM-OBJ-skill-metadata, MEM-OBJ-skill-run, RT-OBJ-agent-step and RT-OBJ-skill-proposal identify deployed instruction, discovery/scope data, retained execution summary, current action payload and proposed revision respectively. RT-RTE-skill-revision changes the deployed procedure after caller-selected apply. Shared skill subject matter creates no further duplicate identity.

The authority channels are distinct. RT-BAP-memory-injection prepends decorator-selected memory to structured-output input; MEM-BAP-session-guidance supplies session guidance after history/question/context; RT-BAP-skill-guidance supplies procedure tool results; MEM-BAP-ranking-state changes code selection. MEM-RTE-session-guidance updates/selects guidance, rather than duplicating RT-RTE-decorated-agent or the whole RT-RTE-recall-dispatch. No channel supersession is warranted.

MEM-CLM-evaluation-boundary covers direct-LLM golden-answer scoring and its unestablished production linkage. EPI-RTE-benchmark-evaluation covers exact-match/GEval execution, persistence and reporting, with EPI-OBJ-eval-metrics as payload. The claim, route and payload are not identical, and the inspected adapters differ. EPI-CLM-durable-lessons is a reusable-knowledge claim, not another declaration of the host start policy RT-CLM-auto-admission.

Integration-issue dispositions and convergence

  1. Fatal QA policy versus per-window outcomes. RT-RTE-session-improve gives the orchestration condition stage.fatal with stage status errored. MEM-RTE-trace-promotion reports that some non-budget per-window errors are continued without propagation and failed windows retain watermarks. These concern different enforcement points. They do not disagree about the fatal-stage branch. The unqualified phrase “fatal QA persistence failure aborts” must not be used to infer every failed window produces an errored stage, or that completion means all QA was admitted. End-to-end aggregation remains the memory report's inspection limit.
  2. Typed branches. MEM-RTE-typed-admission is an alternative to ordinary remember progression: session entries return after writes; SkillRuns are graph-backed and can request proposals. It is not a duplicate of RT-RTE-session-improve or RT-RTE-skill-revision. The runtime describes ordinary session/permanent routes, not universal continuation for every typed branch. Uniform automatic improvement and transport parity remain unsupported.
  3. Progressive disclosure. Memory's annotation on RT-BAP-skill-guidance, epistemic's added load_skill passage, and WF-RTE-skill-loading independently establish the 12,000-character handler cap followed by the loop's 8,000-character result cap. They resolve the runtime's explicitly uninspected handler transformation without contradicting its valid delivery identity. Full stored procedure reception, suffix reload and total token-window bounds are not established.
  4. Agent-profile guidance. MEM-OBJ-session-context and MEM-RTE-trace-promotion retain agent-profile lessons. Memory's read-back account says the inspected agentic builder defaults to QA profile. RT-RTE-agentic-tools and workflow's handoff account name optional active guidance without claiming agent-profile injection. The accounts therefore coexist; distillation/explicit profile selection must not be collapsed into automatic agentic delivery.
  5. Trace windows. MEM-RTE-trace-promotion distinguishes oldest-pending graph persistence from newest-at-most-40 semantic extraction, with a backlog countercase and incomplete individual source links. Nothing in the supplied runtime or workflow requires all traces to be semantically extracted. MEM-OBJ-skill-run and workflow's terminal retention account converge separately on selective tool/result summaries, not full replay or authenticated experience.
  6. Truth recovery. MEM-RTE-truth-maintenance, EPI-RTE-truth-reranking and EPI-CLM-epoch-recovery converge on the implemented epoch guard and its abandoned-basis/reused-epoch countercase. The discrepancy is between a source's strong recovery explanation and its documented countercase, not between analyst reports. It prevents an unconditional coordinate-consistency guarantee; it is not an observed incident or evidence of factual testing.
  7. Withdrawal. MEM-RTE-memory-withdrawal adds a distinct lifecycle route, not a supersession of ingestion or recall. Recorded dependency cleanup coexists with trace-ID gaps, append-only alternatives, unresolved derivative retraction and unknown cache-prune tenant extent. The other members do not require universal erasure or inverse ranking updates. These remain bounded lifecycle limits rather than established system-wide absences.
  8. Evaluator linkage. MEM-CLM-evaluation-boundary, EPI-RTE-benchmark-evaluation, MEM-OBJ-skill-run and workflow's terminal account agree that benchmark comparison is not established as a production install gate. Runtime leaves score provenance uninspected; specialist findings about neutral unscored defaults do not invalidate the score-based selector. No observed improvement, causal attribution or performance acceptance follows.
  9. Remaining overlaps. Memory's no-known-duplicate disposition holds for its comparison against the supplied runtime. Comparison across all four supplied reports adds only the two object supersessions above. Valid containers and material parts remain declared; no further identical referent or required missing part was established from these inputs.

Workflow's inherited evidence-append no-op qualifies terminal delivery in RT-RTE-agentic-tools, rather than contradicting it: runtime promises an answer return, not a final evidence-validation gate. A nonblank answer branch, caught dispatcher errors and best-effort retention are separate from correctness, rollback and host adoption. The no-op finding is supported in the workflow account; it does not license a negative about every other evidence-bearing search route.

RT-RTE-skill-revision's separate theory and property statuses remain unchanged. Memory, epistemic and workflow add lineage and delivery limits without claiming performed content-directed criticism or successful capacity improvement. Afforded reflection/self-improvement can coexist with uninspected learning, autonomous/reflective theory building and actual content-dependent decisions. None is promoted by proposal retention, installation or skill opening.

No unresolved cross-report conflict is established by the supplied findings. Unknown operation, partial coverage, unresolved source/score provenance and bounded recovery/withdrawal limits remain where their declaring reports place them. This statement does not settle support questions reserved for report verification.