Thalo memory profile

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

Comparison rationale

The profile separates retained content from access controls. Entry prose and prior generated Markdown supply knowledge context; retained synthesis prompts supply instructions. Queries, metadata and checkpoints select scope and enforce membership. Schemas validate structural conformance, with enforced CLI failure only when checking is called. None of those forces establishes semantic acceptance or observed model reliance. See MEM-OBJ-instance-entries, MEM-RTE-schema-resolution, RT-BAP-synthesis-guidance and RT-BAP-checkpoint-selection.

Requested query and actualization output are pull. The shipped host then supplies selected entries, stored guidance and the entire previous file to a fresh model invocation without a model retrieval request: that operation is push. An upstream file path alone does not establish identifier-targeted prose selection; full-file supply is coarse. Exact source entity/tag/link matching is a distinct identity selector; freshness filtering is a separate coarse eligibility operation. Neither is inferred relevance ranking. See MEM-RTE-query-readout, RT-RTE-actualize-selection, RT-RTE-pi-synthesis and EPI-RTE-prose-retention.

Curation values follow implemented transformations over retained memory. Timestamp selection suppresses returned instances sharing a composite key without merging or removing retained source content; key collisions can suppress distinct entries. That access filter establishes no curation value, including dedup. Schema alterations evolve resolved constraints; host replacement evolves a retained synthesis artifact. The generated content's semantic relation remains unresolved, so neither the word synthesis nor a requested guide establishes a new claim or reduced claim representation. Physical merge deletion does not establish a history-retaining invalidation mechanism or memory decay. See RT-RTE-actualize-selection, MEM-RTE-schema-resolution, EPI-RTE-file-reconstruction and EPI-RTE-merge-disposition.

Manual and automatic agency are separate admission mechanisms. CI initiation does not turn the Pi writer or publication into human content approval. Human edits and later review are afforded decisions, while generic external caller identity leaves their control unresolved. Automatic compilation and durable recurrent prose generation do not establish trace learning: source entries have caller-dependent experiential provenance. Without bounded absence records neither a trace-learning no nor trace-source inapplicability is warranted. See RT-RTE-local-authoring, RT-RTE-action-publication, MEM-OBJ-instance-entries and RT-RTE-pi-synthesis.

The supplied reconciliation supersedes EPI-OBJ-synthesis-prose with MEM-OBJ-synthesis-prose, and EPI-RTE-query-selection with MEM-RTE-query-readout. All classifications use those canonical records. The Pi guard, reconstruction and retention parts remain valid alongside their runtime container; no value is strengthened by supersession.

All positive findings retain their local basis. Implementation is not observed operation or causal benefit. Partial axes retain unresolved initial provenance, external admissions/consumers, browser later-memory status and provider alternatives rather than obtaining known coverage by dropping them. The parameter component's parametric form is known, but provider retention/update routes remain uninspected. See MEM-OBJ-playground-state and RT-CMP-gpt5-mini.

Source-read log: no source reads were made. Classification used only the supplied accepted members and reconciliation; no records, evidence or quotations were added.