Supermemory profile verification
Type: agentic-system-analyses/types/agentic-system-verification.md
Verification
I checked every axis, its findings and evidence bases, named classification units, unresolved alternatives and inventory assessment in the memory profile. The account supports bounded comparison findings, not a complete inventory or observed performance. All ten axis assessments remain partial. Known units classify their named parts or paths at the stated evidence layer; they do not establish exercised behavior or resolve the opaque engine.
| Axis | Support and disposition |
|---|---|
| Storage substrate | Remote document API objects, the local LRU and Durable Object access state support the service-object and in-memory findings. Graph/vector organization remains claimed, with physical engine storage unresolved. Logical Claude paths do not establish filesystem storage. The grouped service-object unit retains separate warrants for exact files and declared documents/profiles. See MEM-OBJ-claude-file, MEM-OBJ-document-content, MEM-OBJ-profile-summary, RT-OBJ-memory-cache and RT-OBJ-space-state. |
| Representational form | Text, fixed-rule identity/eligibility state and dense numerical representations are distinguished rather than classifying a mixed container from its display. The parametric finding concerns declared embeddings, not an inferred update to model weights. Arbitrary file content and alternative payload transformations retain uncertainty. See MEM-OBJ-memory-state, MEM-OBJ-claude-file, MEM-OBJ-document-content and RT-CMP-local-embeddings. |
| Lineage | Acquisition, file authorship, session extraction and compiled derivatives have distinct derivation paths. The profile does not let a known subsequent transformation resolve initialization or embedding provenance. Trace extraction remains claimed because public submission does not implement the intervening engine. See RT-RTE-vercel-call, MEM-RTE-claude-file, MEM-RTE-engine-dreaming and EPI-RTE-memory-format. |
| Behavioral authority | Findings follow consumed parts and effects: reference text at answering models, interpretive guidance, session experience at the durable-context update, retrieval ranking inputs, namespace routing, eligibility vetoes and upload admission criteria. Instruction is scoped to guidance use, not every arbitrary file. Claimed engine effects are not upgraded to inspected enforcement. See RT-BAP-memory-context, MEM-BAP-file-reload, MEM-RTE-engine-dreaming, MEM-RTE-mcp-read, RT-BAP-space-selection, MEM-BAP-eligibility-selection and RT-RTE-upload-session. |
| Write agency | Automatic capture, model-command admission and declared engine/connector admission are separated from the afforded human guided-save decision. Generic callers and initial population remain unresolved. Starting a request or authoring text is not treated as proof of a human admission decision. See RT-RTE-vercel-call, RT-RTE-tool-loop, MEM-RTE-claude-file, RT-RTE-mcp-save and MEM-RTE-connector-sync. |
| Curation operations | Wired normalization/deduplication and exact-file revision support their local operations. Consolidation, synthesis, evolution, expiry, invalidation and confirmation promotion keep design-level bases. The particular non-entailed graph example supports synthesis without classifying every extraction that way. Approval changes product status, not demonstrated factual warrant. See EPI-RTE-memory-format, MEM-RTE-claude-file, MEM-RTE-engine-dreaming, EPI-RTE-graph-dreaming, MEM-OBJ-profile-summary and MEM-RTE-inference-review. |
| Read-back direction | Explicit tool and prompt requests support pull. Pre-model middleware supply supports push while the upstream API request remains pull. Connector acquisition alone is not counted as model-facing push. See MEM-RTE-mcp-read, RT-RTE-mcp-context, MEM-RTE-claude-file, RT-RTE-vercel-call and MEM-RTE-connector-sync. |
| Read-back signal | Host invocation triggers profile supply; mode/availability supplies the coarse component and namespace selects the profile population. A continuation's actual compound-key match selects cached text. Query-driven recall retains uncertainty about remote semantic scoring rather than inferring an embedding selector from embedding storage. Requested file IDs are not counted as push. See RT-RTE-vercel-call, RT-OBJ-memory-cache and MEM-RTE-livekit-session. |
| Trace learning | The claimed positive spans automatic session capture, durable fact/profile production and later context consumption. Its basis stays claimed because the intermediate transformation is opaque. Arbitrary saves, connector inputs and other updates remain unresolved; no unsupported local or system-wide negative is emitted. See RT-RTE-vercel-call, RT-RTE-openai-call, RT-RTE-python-context, MEM-RTE-livekit-session, MEM-RTE-engine-dreaming and MEM-OBJ-profile-summary. |
| Trace source | Conversational turns support session-logs for the declared qualifying update. LiveKit event callbacks do not turn those turns into environmental event records. Structured tool inputs, mixed imports and other update sources are separately unresolved rather than being covered by the known conversational unit. See MEM-RTE-livekit-session, RT-RTE-vercel-call, MEM-RTE-engine-dreaming and MEM-RTE-connector-sync. |
I resolved the supplied reconciliation's file-object, file-route and review-route supersessions to the memory declarations. The profile uses those canonical citations and preserves arbitrary-content and acceptance qualifications. The broader epistemic factual population and narrower graph transformation remain valid container/part accounts; neither replaces the complete document-processing route. No unresolved inter-report conflict changes an emitted comparison value.
Some units group mechanisms sharing a classification, but their notes and finding-level citations preserve the relevant parts, consumers and evidence layers. They do not transfer a wired witness to a claimed alternative. Uninspected adapter transformations and opaque engine alternatives remain included; their uncertainty prevents complete-value comparisons. See MEM-CLM-adapter-variants and MEM-RTE-engine-dreaming. These are faithfully expressed uncertainties, not additional publication limits. No unsupported value, evidence upgrade, absence or complete-coverage assertion meets the blocker threshold.
Source-read log: none. This judgment used the handed records and reconciliation only; it adds no source evidence or replacement profile text.
Blockers
none
Limits
none