Supermemory reconciliation

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

Reconciliation

Boundary and connected stages

The four reports share the partial product-loop boundary. Public integrations and callers supply implementation evidence; the non-public extraction and maintenance engine supplies declared contracts, not inspected implementation. RT-CLM-engine-boundary retains the source passage establishing that distinction. The Source register contains no observed run or causal experiment. No connection below promotes wiring to operation or benefit, or selects one report's status over another's.

RT-RTE-vercel-call submits conversation text; MEM-RTE-engine-dreaming describes the subsequent remote document-to-chunk/fact/profile chain. These are successive routes, not duplicates. MEM-OBJ-document-content, MEM-OBJ-memory-facts, MEM-OBJ-profile-summary and RT-OBJ-memory-cache distinguish imported content, extracted assertions, profile synthesis and wrapper-local display reuse. The current epistemic ledger assigns document content to Vercel acquisition, and its bounded conclusion explicitly retains the opaque intervening transformation. The memory account's mixed user/generated transcript lineage and the epistemic account's unknown source warrant therefore connect without identifying submission with derived-fact production.

The workflow account selects the explicit OpenAI playground driver within RT-RTE-tool-loop; it declares no duplicate route. Its annotation supplies concurrent batch ordering, trace-assembly failure, deadline and HTTP-delivery details. Under SRC-1, its passages from packages/tools/src/openai/tools.ts and apps/sdk-playground/src/lib/chat-handlers.ts distinguish executor error returns from end-to-end recovery. These refinements do not contradict the runtime's bounded action loop. Neither report establishes an ordered save-then-read transaction, compensating rollback or answer-quality acceptance.

Duplicate supersessions

Amendment: EPI-OBJ-claude-files is superseded by MEM-OBJ-claude-file. Both refer to namespace-scoped remote documents carrying Claude file paths and source/path metadata, not local filesystem files. Their SRC-1 evidence is packages/tools/src/claude-memory.ts. The epistemic arbitrary-content qualification and the memory account's authored-prose and reasons-retention limits remain findings about this same object. Identity establishes neither truth-aptness nor useful rationale for a particular file instance.

Amendment: EPI-RTE-claude-files is superseded by MEM-RTE-claude-file. Both describe the command dispatcher for exact-file view/create/edit/delete/rename in SRC-1, packages/tools/src/claude-memory.ts. Both retain update-before-add replacement and non-atomic destination-write/source-delete rename. The memory record additionally traces the shipped return-to-Claude consumer; the epistemic record adds literal-replacement analysis and the absence of an established content-acceptance criterion. These are compatible findings about one route. Both retain wired implementation and uninspected operation; supersession changes neither status nor evidence scope.

Amendment: EPI-RTE-inference-review is superseded by MEM-RTE-inference-review. Both identify the same namespace-scoped inferred-memory queue and approve/decline/undo transitions, with the same action-table passage from SRC-1, apps/docs/recall/memory-review.mdx. The epistemic record additionally cites the typed interface in packages/lib/api.ts, not endpoint implementation. Memory's claimed design and afforded consumer coexist with epistemic doctrine-only architecture and uninspected implementation. The epistemic ledger and claim comparison retain their finding that confirmation is not demonstrated evidence-consuming acceptance. Neither report describes mandatory pre-search quarantine.

These amendments state identity only. Both declarations remain, and their differing evidence qualifications and analyst-owned findings are not rewritten or discarded.

Containers, parts and ruled-out identities

EPI-OBJ-memory-facts spans graph/profile/search assertions and symbolic eligibility/relationship metadata. It is not identical to the text-only MEM-OBJ-memory-facts, the symbolic MEM-OBJ-memory-state, or the derived sections and aggregates in MEM-OBJ-profile-summary. The cited SRC-1 passages in apps/docs/concepts/graph-memory.mdx, apps/docs/recall/memory-review.mdx and apps/docs/recall/user-profiles.mdx distinguish their content and consumers. The broader epistemic population remains a valid container alongside the separately described parts; no split supersession is required. Its statement that no supplied object names the population does not erase the memory report's narrower declarations. No single part inherits the others' lineage, checks or authority.

EPI-RTE-graph-dreaming addresses graph transformation within the broader indexing/dreaming chain of MEM-RTE-engine-dreaming. The latter also covers chunks and embeddings before graph readiness. Whole-route identity is ruled out. The epistemic record's PM/payment-API example from SRC-1, apps/docs/concepts/graph-memory.mdx, supports its ampliative classification of that example, not every extraction. The memory account's unknown backend algorithms and the epistemic account's otherwise indeterminate extraction relation are compatible. The valid broader route need not be superseded by its narrower graph account.

EPI-RTE-memory-format explicitly declares itself part of RT-RTE-vercel-call. Its default-string projection and normalized deduplication converge with the memory annotation on that runtime route, using SRC-1, packages/tools/src/tools-shared.ts and shared retrieval/prompt helpers. A formatting stage is not the complete retrieval/generation/submission route. EPI-RTE-mutation-checks similarly declares itself part of RT-RTE-tool-loop: helper target/status/count checks are not the model/executor loop. The memory withdrawal annotation supplies compatible detail rather than another declaration. Neither part calls for supersession.

EPI-OBJ-openai-context explicitly declares itself part of RT-RTE-openai-call. It names the transient request payload, not another invocation route. Identity with RT-OBJ-memory-cache is ruled out: the latter is Vercel's in-process LRU, whereas SRC-1, packages/tools/src/openai/middleware.ts, supplies the epistemic record's passages for Chat messages and Responses instructions. Different preparation objects can carry similar assertions without sharing storage or invalidation.

EPI-RTE-search-projection concerns the search-output stage, and EPI-OBJ-search-results its transient result. MEM-RTE-mcp-read covers the broader explicit search/profile/list/document-read family. Exact route identity is ruled out; the projection and output are valid narrower accounts, not replacements for the family. Their SRC-1 evidence in apps/mcp/src/server/tools/search-memory.ts and apps/mcp/src/server/client/index.ts converges on text, ID, similarity and optional document pointers. Search dropping inference status is compatible with list mapping retaining flags. RT-RTE-mcp-context is a separate prompt-request route with different output budgets, not a duplicate of explicit tools.

MEM-RTE-livekit-session is distinct from RT-RTE-python-context, whose principal referent is Agent Framework and whose LiveKit note establishes only hook dispatch. Memory's SRC-1 passages in packages/livekit-sdk-python/src/supermemory_livekit/memory.py add session capture, buffer/retry and assistant-role recall. MEM-CLM-adapter-variants likewise adds selected Mastra/VoltAgent distinctions rather than replacing the Vercel route. MEM-CMP-claude-model and RT-CMP-openai-model name different provider endpoints and callers, not duplicate components. The extraction and embedding components remain separate from those generation endpoints.

MEM-RTE-source-maintenance, MEM-RTE-semantic-forget, the review route, RT-RTE-mcp-forget and framework mutation checks do not name interchangeable withdrawals. Their passages distinguish permanent source deletion, soft forgetting, reversible review disposition, MCP exact/candidate branching and framework normalized-exact target checks. MEM-RTE-connector-sync is continuing external acquisition with resource selection and optional preservation at disconnect; RT-RTE-upload-session is a single-consumption upload capability. Identity is ruled out for these mechanisms despite shared downstream memory consumers.

MEM-BAP-file-reload, MEM-BAP-eligibility-selection, RT-BAP-memory-context and RT-BAP-space-selection retain distinct consumers/channels: exact file prose, service eligibility state, model-facing profile/search text and durable namespace routing. Eventual influence on behavior is not identity. The memory annotation extends text-channel coverage to LiveKit's assistant-role injection without making it system-role enforcement.

MEM-CLM-graph-display, EPI-CLM-graph-truth and RT-CLM-engine-boundary concern rendering, a product warrant claim and unavailable implementation, respectively. They are not duplicate claims. EPI-CLM-review-facts states a confirmation claim, not a second review route. EPI-CLM-benchmark-judge names external answer evaluation, distinct from publication admission in RT-RTE-publish-tools. Neither evaluator is established as a live product memory-truth oracle.

Disagreements and integration dispositions

No current inter-report disagreement is established by these pinned findings. In particular, the earlier OpenAI/cache association is not carried forward: the current epistemic ledger assigns EPI-OBJ-openai-context to RT-RTE-openai-call, while assigning RT-OBJ-memory-cache to Vercel retention. The new object's source passages and the bounded conclusion distinguish transient preparation from Vercel cache reuse. The previously prevented cross-adapter distinction is now explicit; this is not evidence that OpenAI uses the LRU. No correction answers or declined objections were supplied, so no author response is presumed.

The memory integration issues request no correction to the selected runtime trace and propose no remaining runtime duplicate. The supersessions above instead connect cross-specialist declarations. The following issues remain limits, not conflicts between reports:

  • Version skew: MEM-RTE-livekit-session retains legacy signatures alongside the current profile contract in MEM-OBJ-profile-summary. Wiring does not settle deployed SDK/service compatibility. No report asserts adapter-wide equivalence.
  • Readiness and replacement: MEM-RTE-engine-dreaming distinguishes document done from graph readiness; the Claude route distinguishes append-on-add from update replacement. These qualify completion and stable-ID semantics without contradicting runtime's remote-completion limits.
  • Lineage at delivery: memory's read/profile/state records, epistemic formatting/search records and RT-BAP-memory-context independently distinguish available storage metadata from actual model payloads. Optional document reload is compatible with default metadata loss, not automatic grounding. Profile aggregation need not map to one individually removable fact.
  • Withdrawal propagation: MEM-RTE-source-maintenance, MEM-RTE-semantic-forget, MEM-BAP-eligibility-selection and the Vercel cache concern different layers. Remote search exclusion does not establish immediate refresh of aggregates, shared descendants, cached strings or prior model turns. The workflow's lack of rollback remains compatible with these limits.
  • Connector schedules: MEM-RTE-connector-sync retains inconsistent schedule descriptions within SRC-1, apps/docs/connectors/overview.mdx. This is source-document ambiguity, not conflicting report findings. It prevents a uniform timing guarantee and leaves selection-change removal unresolved.
  • Coverage and evaluation: both specialist inventories retain available but incompletely traced alternative integrations and tests. The non-public engine and external benchmark/plugin implementations are access limits under the boundary, not available-file omissions. No member declares an evidenced absence converting those limits into a system-complete negative.

The reports converge on operational admission being distinct from factual acceptance: save, upload, equality/status checks, review labels, publication and HTTP delivery have different targets and force. Review can change search eligibility before demonstrated evidence-consuming acceptance. The benchmark's reference-answer scope does not license claims about extraction mechanisms or component effect. Retention and later delivery likewise do not establish reliable compliance, improved capacity, reflection, theory-building or exercised self-improvement; the members leave those properties separately unestablished within their assessed routes. These are shared bounded conclusions, not one global verdict. This reconciliation used the supplied reports and made no source rereads or target executions.