ai-memex-cli memory profile

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

Comparison rationale

The local artifact inventory separates bodies, access fields and guidance rather than assigning the entire runtime one storage or authority value. MEM-OBJ-page-metadata supports both a files-backed header and a recomputed in-memory access structure. The Git-backed product description does not establish repository-backed memory writes. External model weights and qmd internals are excluded dependencies, not silently classified from their interfaces.

The reconciliation supplies no supersessions. Its valid object containers remain alongside operative parts; this profile cites the narrower memory parts where their consumers differ. No record is renamed or replaced. Classification rests on the supplied accepted records only; no frozen-source file was read for this job.

The strongest trace-fed path is the implemented request to append session learnings to the live schema, followed by later schema loading: RT-RTE-session-distill, MEM-OBJ-wiki-schema and RT-BAP-schema-prompt. Experience has learning authority at that update consumer and the resulting note has instruction authority at the later maintenance agent. The separate summary-to-later-context path also qualifies without requiring a new claim or observed improvement. Mechanical rendering alone does not establish lesson learning. Tool-origin classification is restricted to selected surviving tool-derived session content described by MEM-OBJ-session-markdown; no complete tool transcript or all-source coverage is asserted.

Curation deliberately retains weaker findings. The record of a model request establishes requested consolidation, revision or duplicate treatment, not an observed successful transformation. The fixed digest renderer independently establishes wired consolidation. No synthesize finding follows from a synthesis page type: EPI-CLM-synthetic-pages and the route assessments leave actual new-claim production unresolved. Template seeding, primary-key refusal and simple index rebuilding are not independently called curation.

Read-back separates requested retrieval from preselected context. RT-RTE-context-delivery supports pull search/injection and a distinct after-clean persistent supply route. The latter is push to host files, not proof of unsolicited delivery into an already running model. Broad digest slicing is coarse selection; exact bound scene names select retained manifest guidance, but registry destination keys do not establish page-level identifier matching. Scene-binding loss and missing per-entry vault identity in MEM-OBJ-context-registry limit the persistence of that targeting without negating its initial wired path.

The partial axes retain unknown initial manifest admission, generic caller control, mixed trace inputs and unseen semantic outputs. Those gaps prevent exhaustive lineage, admission, curation and trace classifications; they do not erase positive paths. Actual user approval, model compliance, host activation, accuracy and benefit remain unobserved throughout. The conflicting foreground-save alternatives in MEM-RTE-query-save are classified separately rather than promoted into a universal human gate.