Supermemory epistemic report

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

Source-and-claim boundary

The question is what licenses reliance on submitted, inferred and retrieved content, and what results can change later behavior. The boundary and Source register establish SRC-1: the public artifact is complete, but the product loop is partial. The non-public extraction, embedding, graph, ranking and maintenance engine prevents an implementation-grounded account of its truth checks. Provider weights and their updates are also inaccessible; see RT-CMP-extraction-model, RT-CMP-local-embeddings and RT-CMP-openai-model. No observed run or causal experiment is registered. Source deductions below are implementation evidence, not target executions.

The consequential warrant claims are EPI-CLM-graph-truth, EPI-CLM-review-facts, EPI-CLM-benchmark-judge and the engine boundary in RT-CLM-engine-boundary. The review endpoints are a material alternative omitted from the provisional runtime account, not a correction to its selected middleware trace. They defeat an unqualified claim that the product declares no confirmation or rejection route. Their existence in documentation does not establish a working human gate on every read or write.

Entry point or operation	Source path, SRC-1	Coverage and evidence basis	Remaining limit
Automatic Vercel and OpenAI context/save	packages/tools/src/vercel/index.ts; packages/tools/src/openai/middleware.ts; packages/tools/src/conversations-client.ts; packages/tools/src/vercel/middleware.ts	Supplied RT-RTE-vercel-call, RT-RTE-openai-call; direct reads of OpenAI preparation, Vercel conversion and conversation serialization	Provider compliance and remote admission unobserved; OpenAI prepared context is not the Vercel LRU
Shared retrieval, deduplication and prompt construction	packages/tools/src/shared/memory-client.ts; packages/tools/src/tools-shared.ts; packages/tools/src/shared/prompt-builder.ts; packages/tools/src/shared/memory-context.ts	Direct implementation reads; EPI-RTE-memory-format	Custom templates can differ; no universal semantic preservation guarantee
MCP dispatch, save form and search	apps/mcp/src/server/tools/index.ts; apps/mcp/src/server/tools/guided-save.ts; apps/mcp/src/server/tools/search-memory.ts; apps/mcp/src/server/client/index.ts	Direct registry, form and search reads; client mapping section; EPI-RTE-search-projection. Supplied RT-RTE-mcp-save	Form affords review, not mandatory approval; graph/list/profile/document handlers not independently audited here
MCP forget, context and upload	apps/mcp/src/server/tools/forget-memory.ts; apps/mcp/src/server/prompts/context.ts; apps/mcp/src/server/space-state.ts	Supplied RT-RTE-mcp-forget, RT-RTE-mcp-context, RT-RTE-upload-session	Backend withdrawal and host-history eviction unverified
Framework tools, deletion and forget helpers	packages/tools/src/ai-sdk.ts; packages/tools/src/tools-shared.ts	Supplied RT-RTE-tool-loop; direct helper read; EPI-RTE-mutation-checks	Request checks are not content-truth checks
Claude memory command dispatch and all file operations	packages/tools/src/claude-memory.ts	Direct complete implementation read; EPI-RTE-claude-files	Remote replacement, concurrent writes and rename atomicity unverified
Inference review queue/actions	apps/docs/recall/memory-review.mdx; packages/lib/api.ts	Direct documentation and schema-section reads; EPI-RTE-inference-review	No inspected endpoint implementation or exercised review
Extraction, derives, updates and forgetting	apps/docs/concepts/graph-memory.mdx	Direct doctrine read; EPI-RTE-graph-dreaming	Non-public engine prevents implementation judgment
Agent Framework lifecycle, LiveKit and release gate	packages/agent-framework-python/src/supermemory_agent_framework/context_provider.py; packages/livekit-sdk-python/src/supermemory_livekit/agent.py; .github/workflows/publish-tools.yml	Supplied RT-RTE-python-context, RT-RTE-publish-tools	Other Python, Mastra and Voltagent internals uninspected; no adapter-wide negative
Benchmark evaluator	apps/docs/memorybench/memscore.mdx	Direct doctrine read; EPI-CLM-benchmark-judge	External benchmark implementation, runs and comparisons unavailable
Other shipped relevant paths	packages/memory-graph/; apps/memory-graph-playground/; apps/docs/connectors/overview.mdx; skills/supermemory/SKILL.md	Available but not independently traced here beyond supplied boundary coverage	No system-complete negative about evaluation, lineage or maintenance; visualization and connector behavior not assessed

Epistemic-object inventory

Fields owned by records are referenced rather than restated.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-memory-facts	Personal facts, preferences and inferred propositions	see record	SRC-1, apps/docs/concepts/graph-memory.mdx	Actual provenance and warrant unknown
RT-OBJ-memory-cache	Formatted retrieved assertions	see record	SRC-1, packages/tools/src/shared/cache.ts	Vercel-local retention is not acceptance
EPI-OBJ-openai-context	Prepared retrieved assertions	see record	SRC-1, packages/tools/src/openai/middleware.ts	Transient preparation, not a retained-cache path
MEM-OBJ-document-content	Submitted conversation assertions	see record	SRC-1, packages/tools/src/conversations-client.ts	Raw input, not derived graph facts
EPI-OBJ-search-results	Selected assertions or document chunks	see record	SRC-1, apps/mcp/src/server/tools/search-memory.ts	Inference distinction lost on this search path
EPI-OBJ-claude-files	Arbitrary file text; truth-aptness not determinable without an instance	see record	SRC-1, packages/tools/src/claude-memory.ts	File mutations can change claims without assessing them
RT-OBJ-space-state	None: routing/capability state	see record	SRC-1, apps/mcp/src/server/space-state.ts	Operational scope selection, not warrant

Authority-route ledger

The architectural status assesses each route, separately from operation evidence. Every implementation route below remains unobserved. References in cells delegate the detailed field to its declaring record.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-vercel-call	content transformation	implemented	MEM-OBJ-document-content	truth-apt transformation: acquisition/import	Serialized conversation enters document-ingestion interface	see route	see route	see route	API submission request	Imported conversation assertions; source warrant unknown	Save eligible text, including generated text	see route	SRC-1, packages/tools/src/conversations-client.ts	none	none	Conversion omits system text and selected payloads; acquisition does not classify remote document-to-fact derivation
RT-RTE-openai-call	operational admission/selection/consumption	implemented	EPI-OBJ-openai-context	no content change	Prepared context delivered in Chat messages or Responses instructions	see route	see route	see route	Model-input preparation	No new factual warrant	see route	see RT-BAP-memory-context	SRC-1, packages/tools/src/openai/middleware.ts	none	none	Fail-open continuation is not endorsement; no Vercel LRU association
EPI-RTE-memory-format	content transformation	implemented	EPI-OBJ-memory-facts	truth-apt transformation: non-ampliative reshaping	Default textual projection	see route	see route	see route	Formatting/selection	No truth license; epistemic annotations can be lost	see route	see route	SRC-1, packages/tools/src/shared/memory-client.ts	none	none	Arbitrary custom template relation indeterminate
EPI-RTE-search-projection	operational admission/selection/consumption	implemented	EPI-OBJ-search-results	truth-apt transformation: non-ampliative reshaping	API results to tool output	see route	see route	see route	Tool-result delivery	Similarity is not confidence in truth	see route	see route	SRC-1, apps/mcp/src/server/tools/search-memory.ts	EPI-CLM-review-facts	qualification	Search output does not expose inferred status
EPI-RTE-mutation-checks	check/evidence production	implemented	EPI-OBJ-memory-facts	no content change	Mutation target/status and response count	see route	see route	see route	Protocol gates	Equality/status/count only	see route	see route	SRC-1, packages/tools/src/tools-shared.ts	none	none	Cannot establish truth or remote deletion correctness
RT-RTE-mcp-forget	disposition/acceptance	implemented	EPI-OBJ-memory-facts	no content change	Withdrawal selection, not factual acceptance	see route	see route	see route	Dispatch or return candidates	No falsehood judgment follows from forget	see route	see route	SRC-1, apps/mcp/src/server/tools/forget-memory.ts	none	none	No mandatory human gate or local cache eviction
EPI-RTE-claude-files	retention	implemented	EPI-OBJ-claude-files	truth-apt transformation: indeterminate	Caller-selected text changes	see route	see route	see route	Remote write/read requests	No acceptance criterion	see route	see route	SRC-1, packages/tools/src/claude-memory.ts	none	none	Instances may instead be non-truth-apt instructions
RT-RTE-vercel-call	retention	implemented	RT-OBJ-memory-cache	no content change	Continuation cache	see route	see route	see route	Cache reuse	Neither freshness nor endorsement	see route	see route	SRC-1, packages/tools/src/shared/cache.ts	none	none	No backend-mutation invalidation
EPI-RTE-graph-dreaming	content transformation	doctrine only	EPI-OBJ-memory-facts	truth-apt transformation: ampliative conjecture	Derives example	see route	see route	see route	No inspected engine force	Candidate inference, not verified consequence	see route	see route	SRC-1, apps/docs/concepts/graph-memory.mdx	EPI-CLM-graph-truth	implementation unknown	Classification is of documented example, not all extraction
EPI-RTE-inference-review	disposition/acceptance	doctrine only	EPI-OBJ-memory-facts	no content change	Confirm/decline/undo one inferred fact	see route	see route	see route	Declared status/search changes; no inspected enforcement	Confirmation label without supplied evidence criterion	see route	see route	SRC-1, apps/docs/recall/memory-review.mdx	EPI-CLM-review-facts	acceptance limit	No evidence-consuming acceptance or lifecycle integration established
RT-RTE-mcp-context	operational admission/selection/consumption	implemented	RT-OBJ-space-state	no content change	Namespace and fact selection	see route	see route	see route	Routing and bounded prompt construction	Namespace access is not factual warrant	see route	see RT-BAP-space-selection, RT-BAP-memory-context	SRC-1, apps/mcp/src/server/prompts/context.ts	none	none	Host placement and compliance unknown
RT-RTE-upload-session	disposition/acceptance	implemented	RT-OBJ-space-state	no content change	Upload capability admission	see route	see route	see route	Request authorization	No content-quality license	see route	see route	SRC-1, apps/mcp/src/server/space-state.ts	none	none	Single consumption not exactly-once ingestion
RT-RTE-publish-tools	disposition/acceptance	implemented	none	non-truth-apt policy/content update: package publication	Version/build/compatibility gate	see route	see route	see route	Ordered CI publication gate	Compatibility only, not knowledge truth	see route	see route	SRC-1, .github/workflows/publish-tools.yml	none	none	No executed CI evidence

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-graph-truth	see record	SRC-1, apps/docs/concepts/graph-memory.mdx; doctrine/design	see record	No inspected engine route; EPI-RTE-graph-dreaming is doctrine only	none supplied	No comparison or engine execution	Declared extraction and conjecture; correctness unknown	“Stays true” exceeds public implementation evidence
EPI-CLM-review-facts	see record	SRC-1, apps/docs/recall/memory-review.mdx; doctrine/design	see record	Schema support, not endpoint enforcement; EPI-RTE-search-projection shows a consumer	none supplied	Review procedure and factual evidence unknown	Declared reversible disposition; not demonstrated acceptance	Unreviewed suggestions already searchable; inference marker not carried on this MCP search path
EPI-CLM-benchmark-judge	see record	SRC-1, apps/docs/memorybench/memscore.mdx; doctrine/design	see record	none inspected in registered artifact	none supplied	External implementation/run; judge choice, prompts and excluded failures limit interpretation	A described answer evaluator, not a live graph-admission oracle	No supported component-effect or general truth guarantee
RT-CLM-engine-boundary	see record	SRC-1, apps/docs/self-hosting/overview.mdx; doctrine/design	see record	Public routes delegate remotely	none supplied	No engine code	Partial loop remains the right boundary	Implementation questions remain open, not negative findings

Bounded conclusion

Public routes acquire and reuse assertions without establishing their truth. Automatic conversation submission can feed user and generated assistant text into later remote profile/search results; source warrant is unknown. See RT-RTE-vercel-call, RT-RTE-openai-call and RT-RTE-python-context. The acquired input is document content, not already-derived graph assertions; see MEM-OBJ-document-content. The intervening document-to-fact processing is separately declared in MEM-RTE-engine-dreaming, not established by submission code. Vercel cache retention and OpenAI transient preparation are separate paths; see RT-OBJ-memory-cache and EPI-OBJ-openai-context. Neither retention nor repeated delivery licenses acceptance. The backend extraction relation remains indeterminate; no inspected route establishes an entailed new factual proposition from warranted premises.

The default text consumers can weaken epistemic distinctions. EPI-RTE-memory-format drops metadata and merges normalized variants. EPI-RTE-search-projection drops inference status, although the client separately preserves it for memory-list entries. Similarity and profile categories govern supply, not a truth decision. These are consumer-specific findings, not claims that the API never offers lineage. Model-facing text is advisory/interpretive through RT-BAP-memory-context; code enforces replacement of owned blocks, not compliance with their content.

Conjecture and review are declared, but acceptance is not demonstrated. The specific graph example is ampliative because the stated premises also fit work on another product; see EPI-RTE-graph-dreaming. Review declares confirmation, rejection and undo and directly affects future search, yet supplies no required evidence criterion, intended-use test or inspectable implementation. No accepted ampliative claim or post-acceptance lifecycle integration is established by EPI-RTE-inference-review. Unreviewed content may have operational use first. This does not establish that server checks are absent.

Withdrawal and editing have narrower, real protocol targets. Equality/status/count checks in EPI-RTE-mutation-checks and candidate gating in RT-RTE-mcp-forget constrain requests, not factual warrant. Remote forgetting does not evict already injected prompts or local caches. EPI-RTE-claude-files provides inspectable text revision and later exact-file reads without content-quality judgment; rename is write-then-delete, not an atomic epistemic revision. Guided save affords human editing, but direct add remains available.

Evaluation does not transfer beyond its target. The documented external benchmark judge checks answers against reference answers; it does not establish extraction correctness, graph mechanism, replay safety or component effect. The release workflow gates publication, not memory truth. Neither supplies observed benefit here. Retained trace-fed content can reach later behavior, but that alone does not establish improved capacity, reflection, autonomous theory-building or exercised self-improvement. These properties remain separately unestablished, not bundled into a system-wide verdict. Available uninspected adapters, graph consumers and connector paths also prevent a system-complete negative.

Shared records

Operative objects

EPI-OBJ-openai-context

Label: OpenAI prepared memory context

Part of: RT-RTE-openai-call

This object is the transient context payload within the supplied route, not another route. Identity: retrieved/formatted memories inserted into Chat messages or Responses instructions during request preparation. Form: natural-language assertions inside symbolic message/instruction fields and owned delimiters. Substrate: in-process request values, then external provider input; host history retention is uninspected. Distinct from RT-OBJ-memory-cache, the Vercel-only LRU: the OpenAI preparation retrieves/formats and forwards request values, without that cache. SRC-1 implementation, packages/tools/src/openai/middleware.ts. Its provider consumer and interpretive force are described in RT-BAP-memory-context; no factual acceptance follows. No-text branches skip search and caught retrieval failures clear owned context rather than proving current memories were delivered.

enhancedInstructions = replaceMemoryContext( params.instructions || "", typeof memories === "string" ? memories : "", ) --- packages/tools/src/openai/middleware.ts

messages: enhancedMessages, --- packages/tools/src/openai/middleware.ts

EPI-OBJ-memory-facts

Label: Graph memories and profile facts

Identity: remote graph facts and API profile/search assertions, including inferred facts. Form: natural-language assertions with symbolic IDs, inference/review/latest/forgotten flags and relationships; backend graph storage uninspected. These are remote contents, distinct from the local formatted-string container RT-OBJ-memory-cache; no supplied record names this remote factual population. Source lineage is declared but actual claim-to-evidence links and extraction fidelity remain unknown. SRC-1, apps/docs/concepts/graph-memory.mdx, apps/docs/recall/memory-review.mdx.

| What | Raw input you send | Facts Supermemory extracts | --- apps/docs/concepts/graph-memory.mdx

EPI-OBJ-search-results

Label: MCP matching memories

Identity: search_memory result projection: natural-language memory/chunk text with symbolic IDs, document ID when present, similarity, totals and timing; transient tool response, potentially retained by external host history. Distinct from RT-OBJ-memory-cache: different producer, output and consumer, no wrapper-local LRU. SRC-1, apps/mcp/src/server/tools/search-memory.ts, apps/mcp/src/server/client/index.ts. The latter's mapMemoryEntry preserves isInference on the memory-list path, counterevidence to any API-wide claim of marker removal.

const results = searchResult.results.map((result) => ({ id: result.id, text: getMemoryText(result), similarity: result.similarity, ...(result.documentId && { documentId: result.documentId }), })) --- apps/mcp/src/server/tools/search-memory.ts

EPI-OBJ-claude-files

Label: Claude memory file documents

Identity: namespace-scoped remote documents with source: claude-memory, file paths and reversible IDs. Text is natural-language or arbitrary caller content, with symbolic path/metadata; storage is remote service, not a local filesystem. No counterpart in supplied objects: RT-OBJ-space-state is namespace/capability routing state, not these documents. SRC-1, packages/tools/src/claude-memory.ts.

interface ClaudeFileDocument { documentId: string content: string metadata: ClaudeFileMetadata } --- packages/tools/src/claude-memory.ts

Routes

EPI-RTE-memory-format

Label: Default memory text formatting

Part of: RT-RTE-vercel-call

This is the shared-helper content-transformation stage of that call, not a replacement for its invocation/persistence account. Code fetches profile/search, extracts strings, deduplicates by normalized text with static-before-dynamic-before-search precedence, then passes Markdown strings and selected raw results to a template. The default template uses only the strings. SRC-1 implementation: packages/tools/src/tools-shared.ts, packages/tools/src/shared/memory-client.ts, packages/tools/src/shared/prompt-builder.ts, packages/tools/src/shared/memory-context.ts.

Concrete source deduction, not execution: in full mode static Alex works at Stripe and search [2026-10-19] Alex works at Stripe have the same normalization key; the static string survives and the dated search result is omitted. The default result contains ## Static Profile and the surviving bullet, without the search entry's date or metadata. This is non-ampliative textual selection, but not guaranteed preservation of temporal scope or warrant. In query mode the profile is excluded before deduplication, so the search result survives: counterexample to an unconditional profile-precedence claim. A custom template can consume raw result metadata or produce arbitrary text; its semantic relation is indeterminate without its implementation.

• Immediate return: Formatted memory string, or empty string.
• Later read-back: Wrapper can cache the string for a continuation; see parent route for cache policy.
• Delegated visibility: Default model context gets fact strings, not original result metadata, review flags or a proof.
• Selection predicate: Mode, nonempty memory/chunk, normalized text uniqueness and source priority; no truth evaluator or answer oracle.
• Invalidation or expiry: Owned marked blocks replaced on later preparation; cache invalidation is parent-owned, not remote fact revision tracking.
• Activation or effect: Code determines supplied context; downstream model reliance unobserved.
• Evidence limits: No execution; custom templates and remote semantic fidelity uninspected.
• implementation conclusion status: wired
• operation conclusion status: uninspected

.replace(/^[recent]\s/i, "") .replace(/^[\d{4}-\d{2}-\d{2}]\s/, "") --- packages/tools/src/tools-shared.ts

return User Supermemories: \n${data.userMemories}\n${data.generalSearchMemories}.trim() --- packages/tools/src/shared/prompt-builder.ts

EPI-RTE-search-projection

Label: MCP search result projection

Distinct from supplied routes: RT-RTE-tool-loop is the framework executor; RT-RTE-mcp-context is the MCP profile prompt, not this registered search operation. SRC-1 implementation: apps/mcp/src/server/client/index.ts mapping section and apps/mcp/src/server/tools/search-memory.ts. Backend results become text/chunks, IDs, similarity and optional source-document ID. Neither projection checks truth nor carries inference status. Document IDs preserve a possible lookup path but not a demonstrated warrant chain.

• Immediate return: Text and structured matching memories, totals and timing, or tool error.
• Later read-back: uninspected — host can retain tool history; this handler makes no durable write.
• Delegated visibility: MCP caller receives projected text/IDs/similarity, not inferred/review status.
• Selection predicate: Caller query and resolved namespace; remote search plus client projection, no truth oracle.
• Invalidation or expiry: inapplicable — transient output; no local eviction of previous host messages.
• Activation or effect: Supplies tool context, advisory to host/model; model use unobserved.
• Evidence limits: No service execution; other routes such as memory listing preserve more metadata.
• implementation conclusion status: wired
• operation conclusion status: uninspected

const similarity = (result.similarity * 100).toFixed(0) --- apps/mcp/src/server/tools/search-memory.ts

The load-bearing output-field passage is in EPI-OBJ-search-results.

EPI-RTE-mutation-checks

Label: Framework mutation target checks

Part of: RT-RTE-tool-loop

This record owns the inspected shared-helper check stage, not the model's disposition or the remote mutation implementation. In open requests the caller/model proposes a document ID or memory ID/content; code may veto, and the service may reject. deleteDocument fetches the document, rejects a known nonterminal status, requests deletion and checks count one. Missing status is not rejected by that predicate. forgetMemory with content requests a dry-run semantic preview and selects normalized exact matches; direct ID bypasses preview. Post-request nonzero count is treated as success. The MCP forget handler has a different exact-text rule; see RT-RTE-mcp-forget. SRC-1 implementation, packages/tools/src/tools-shared.ts.

• Immediate return: Count/void on checked protocol success, otherwise error.
• Later read-back: Service mutation can affect later retrieval; visibility and durability uninspected.
• Delegated visibility: Service receives IDs; caller gets success/error, not a factual evaluation.
• Selection predicate: Document existence/status, normalized memory equality or explicit ID, then returned deletion/forget count. No answer oracle, improvement trigger or per-content human approval.
• Invalidation or expiry: Remote withdrawal requested; no local restoration or cache eviction wired by helper.
• Activation or effect: Blocks certain target mismatches/nonterminal deletes and rejects reported zero-success results; not proof of remote effect.
• Evidence limits: Status omission and direct-ID branches limit gates; no truth criterion or rollback.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if (status !== undefined && !TERMINAL_DOCUMENT_STATUSES.has(status)) { --- packages/tools/src/tools-shared.ts

.filter((match) => normalizeMemoryFact(match.memory) === target) .map((match) => match.id) --- packages/tools/src/tools-shared.ts

EPI-RTE-claude-files

Label: Claude document file editing

No counterpart in supplied routes: this dispatches file commands, unlike conversation submission or generic framework fact tools in RT-RTE-tool-loop. Caller supplies command and text; code validates path/required arguments and resolves the exact stored document using source/path metadata rather than ranked chunks. Create overwrites; replace changes the first literal match; insert modifies a line position; rename writes destination then deletes source; delete checks returned count. writeFile tries update, falling back to add only on not-found. These are operational file admissions, not evidence-consuming claim acceptance. Guidance is the shipped command schema and path rules; operating mode is open requests, with no answer oracle. Caller can abstain or later edit/delete, code and service can veto, but no per-content human approval or transactional rollback is established. SRC-1 implementation: packages/tools/src/claude-memory.ts.

• Immediate return: Success/content/error or formatted tool-result with is_error.
• Later read-back: View resolves complete remote file and numbers selected lines; directory view paginates all results then filters paths. A later caller can receive retained edits.
• Delegated visibility: API receives file content and metadata; host gets command outcome or numbered content.
• Selection predicate: Namespace, reversible/legacy IDs, source/path match and command-specific literal/line checks; caller chooses content.
• Invalidation or expiry: Replacement/deletion requests change remote documents; rename is non-atomic write then delete; no local cache or host-context clearing.
• Activation or effect: Retention/read requests wired; actual later behavior depends on external caller interpretation.
• Evidence limits: No concrete file instance or execution; truth-aptness and semantic classification indeterminate, service consistency/concurrency unverified.
• implementation conclusion status: wired
• operation conclusion status: uninspected

const newContent = originalContent.replace(oldStr, () => newStr) --- packages/tools/src/claude-memory.ts

Concrete source deduction: old content A A, old string A, replacement $& becomes literal $& A, not A A. This establishes only the fixed string transformation and literal replacement behavior, not a factual derivation.

if (!isNotFoundError(error)) throw error await this.client.add(this.namespace, { content, id, metadata }) --- packages/tools/src/claude-memory.ts

EPI-RTE-graph-dreaming

Label: Dreaming inferred facts

No supplied route describes this remote internal transformation; RT-CMP-extraction-model names a component, not an inspected route. Documentation says dreaming extracts, links, resolves updates and derives propositions from groups of documents. In the specific example, being a PM and frequently discussing payment APIs/fraud does not entail working on the core payments product: those premises also fit a PM on another product. Thus this example is ampliative conjecture. Other extraction could reshape, derive or conjecture; inaccessible implementation leaves that classification indeterminate. SRC-1 doctrine/design: apps/docs/concepts/graph-memory.mdx.

• Immediate return: uninspected — documentation describes asynchronous/dynamic or instant processing, not an inspected execution.
• Later read-back: Declared graph facts and relations enter search/profile; actual effects unobserved.
• Delegated visibility: Remote engine's model prompts, evidence and intermediate judgments unavailable.
• Selection predicate: Declared related-document grouping and configured dreaming mode; caller supplies documents, remote process decides extraction/relationships. No supplied answer oracle or mandatory human gate.
• Invalidation or expiry: Documentation declares updates, time expiry, contradiction resolution and noise filtering; exact conditions/rollback uninspected.
• Activation or effect: Declared production of retained inferred content, not inspected or observed engine operation.
• Evidence limits: Truth tests, evaluator domain and error calibration unknown; novelty or coherence cannot substitute for acceptance.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

Memory 1: "Alex is a PM at Stripe" Memory 2: "Alex frequently discusses payment APIs and fraud detection" → Derived: "Alex likely works on Stripe's core payments product" --- apps/docs/concepts/graph-memory.mdx

EPI-RTE-inference-review

Label: Inferred memory review actions

No counterpart in supplied runtime routes: RT-RTE-mcp-forget selects forgetting targets; these declared endpoints operate on an inferred-memory queue and review stamps. SRC-1 doctrine/design: apps/docs/recall/memory-review.mdx; interface schema in packages/lib/api.ts, not endpoint implementation. Target is one namespace-scoped inferred proposition. The client proposes approve/decline/undo and the declared server transition decides status. Human UI review is intended but generic HTTP calls do not prove human authorship. No required evidence input, reference truth, intended-use criterion or recorded reasoning is specified. Consequently approval licenses the product's confirmation label, not established evidence-consuming acceptance under this analysis.

• Immediate return: Declared ID/inference/forgotten/review status, or authentication/permission/not-reviewable error.
• Later read-back: Declared approval clears inference marker; decline removes from search and queue; undo restores unreviewed state and can un-forget.
• Delegated visibility: Client gets candidate text, parent count and metadata; no required premise passages or proof supplied by the queue contract.
• Selection predicate: Declared unreviewed, nonforgotten, nonexpired inferred facts; up to 50 ordered by parent count then creation time. Client chooses action; no answer oracle or improvement trigger.
• Invalidation or expiry: Undo reverses prior disposition; decline withdraws remote search availability by contract, not existing host contexts/caches.
• Activation or effect: Declared search eligibility and inference-flag changes; typed action interface exists, but service and consumer execution unobserved.
• Evidence limits: Parent count is not independent corroboration or a truth probability; implementation, human judgment and evidence-consuming acceptance uninspected.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

| Approve | isInference cleared | Stays in search, no longer flagged as inferred | | Decline | isForgotten set | Removed from search entirely — a rejected guess is forgotten | | Undo | back to unreviewed | Returns to the queue, flagged as inferred again | --- apps/docs/recall/memory-review.mdx

Claims

EPI-CLM-graph-truth

Label: Graph stays true

Claimed operation: living graph extraction, updates, extensions, derives, expiry and contradiction handling keep understanding current. Distinct from RT-CLM-engine-boundary, which locates unavailable implementation, not the asserted knowledge mechanism. SRC-1 doctrine/design, apps/docs/concepts/graph-memory.mdx.

How understanding is stored and stays true over time. --- apps/docs/concepts/graph-memory.mdx

• Contradiction — updates win for “what’s true now.” --- apps/docs/concepts/graph-memory.mdx

EPI-CLM-review-facts

Label: Review confirms inferred facts

Claimed operation: approve promotes an inferred memory to confirmed fact, decline forgets it and undo restores unreviewed status. Distinct from RT-CLM-engine-boundary: this is a disposition claim with declared consumers. SRC-1 doctrine/design, apps/docs/recall/memory-review.mdx; schema existence in packages/lib/api.ts does not establish enforcement.

• approve — Promote the memory: clears isInference, so it's treated as a confirmed fact. Stamps reviewStatus: "approved". --- apps/docs/recall/memory-review.mdx

EPI-CLM-benchmark-judge

Label: MemoryBench answer correctness

Claimed operation: an external judge LLM compares benchmark provider answers to ground truth and reports verdicts, with latency and context cost separately reported. No supplied claim counterpart after comparison with RT-CLM-engine-boundary. SRC-1 doctrine/design, apps/docs/memorybench/memscore.mdx. Target/domain: answers to selected benchmark questions under selected judging prompts, not all retained memories. Evaluator choices include GPT-4o, Claude Sonnet and Gemini Flash; their exact versions, weights, update behavior and endpoint resolution are uninspected because no benchmark implementation or execution is registered. No causal experiment or wired product-consumer path is established.

Correctness isn't decided by string matching — a judge LLM (GPT-4o, Claude Sonnet, or Gemini Flash, your choice) compares the provider's answer against ground truth and returns a verdict: --- apps/docs/memorybench/memscore.mdx

| Success rate | Percentage of questions that completed without erroring (failures are excluded from accuracy, not counted against it) | --- apps/docs/memorybench/memscore.mdx