IsaacCLupus_mnemosyn_spec memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

All ten axes remain partial. This is an inventory of proposed memory and the legacy adapter's retained outputs, not a claim that the unified OS runs. Claimed design, afforded external consumption and wired copying stay separate. Missing execution evidence prevents observed or causal findings, not the wired local copying classification. The unresolved backend boundary is supported by RT-RTE-client-query, RT-RTE-link-remember and MEM-OBJ-state-metadata.

Canonical identities and scope

Reconciliation supersedes EPI-OBJ-unified-pages with MEM-OBJ-unified-vault, EPI-OBJ-user-profiles with MEM-OBJ-user-profiles, and EPI-RTE-profile-promotion with MEM-RTE-conversation-promotion. Those canonical objects/routes supply the corresponding storage, form, lineage, admission and trace classifications. Supersession changes citation identity, not evidence strength. The article functions, feedback subroute, graph context and exports remain distinct parts/routes rather than replacements for RT-RTE-design-lifecycle.

The memory boundary excludes the shipped compilation template itself and current-request proxy queues. It includes the use-produced source/related-page material rendered through that template and any provider-retained alternative whose contents remain unresolved. Static template instructions do not establish memory instruction authority; use-retained criticism does, through MEM-BAP-feedback-guidance. Model aliases establish neither a vector substrate nor a use-fed parametric-memory update: see RT-RTE-model-proxy and MEM-OBJ-state-metadata.

Mapping limits

Legacy publication imports page bodies and compiles selected metadata; repeated overwrites evolve existing entries. It is not semantic consolidation or deduplication merely because it normalizes metadata, renames collisions and rebuilds indexes. Optional Git history is a conditional wired repository route, not observed recovery. Its initial source authorship, Link reload payload and downstream ranking remain unresolved: RT-RTE-vault-publish, MEM-OBJ-unified-vault, EPI-RTE-maturity-mapping.

Conversation extraction consolidates retained turns; daily heat decay, hot-tier promotion and archival have separate proposed effects. Profile visibility is not settled. The committed-memory branch supplies a claimed trace-learning chain without requiring automatic final approval or proving improvement. The feedback chain separately retains audit execution feedback, gives it instruction and learning force at a later rewrite, and yields proposed durable article updates. These classifications do not assert accepted new knowledge, parameter training, theory building or exercised self-improvement: MEM-RTE-conversation-promotion, MEM-RTE-feedback-recompile, MEM-OBJ-user-profiles.

Read-back operations are distinguished within a chain. A requested backend response is pull. The graph study's get_context request is also pull, while its automatic addition of selected context to the answering model is push. The page-ID/node match supplies identifier targeting; first-five/edge-availability filtering supplies a separate coarse constraint. No semantic ranking is inferred. Feedback insertion supports proposed push but its inconsistent lookup does not establish a complete identifier selector. Related-page truncation supports coarse filtering without resolving upstream selection: RT-RTE-client-query, MEM-RTE-graph-context, MEM-RTE-feedback-recompile, RT-OBJ-compile-template. The documented file consumer remains afforded pull, not an observed import: MEM-RTE-pack-export.

No negative trace-learning finding is supplied: the accepted records declare no bounded absence warrant for these unresolved alternatives. Likewise, pull-only units do not make the whole read-back-signal axis inapplicable. Complete statistics cannot be derived from these partial inventories, even where a local wired witness establishes a value's existence.

Source-read log

No source files were read for this profile. Classification uses only the supplied accepted records and their annotations, with identity resolution from the supplied reconciliation. No new evidence, quotations or records are introduced.