nao memory profile

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

Comparison rationale

Boundary and evidence strength

All axes are partial inventories, not complete-value assertions. Inspected software paths establish local wiring; the accepted set supplies neither runs nor causal experiments. Configured file-provider alternatives and human-facing maintenance/fix surfaces retain their afforded basis where their full consumer or provider implementation is uninspected. Missing behavior evidence does not make a known architecture unknown, and a known architecture does not establish activation or benefit.

The scope separates accumulated content from the whole runtime. Initial pretrained weights and static shipped instructions are not positive memory values. Opaque provider retention/adaptation remains an unresolved alternative. Unconsumed credentials and transient ToolContext grants do not become memory controls merely because code uses them. The retained evidence-valued part is EPI-OBJ-query-results, not every field of RT-OBJ-tool-context.

Content, control and distinct consumers

User-memory prose supplies knowledge for profile statements and instruction for standing rules through MEM-BAP-user-memory. Its IDs and supersession fields instead supply code-level routing, admission criteria and exclusion. Content category/order/budget filtering is coarse; actual user/source-chat matching is identifier selection. The same retained experience has learning authority at its update consumer, independently of its output's downstream authority. These paths are preserved in MEM-RTE-memory-extraction and MEM-OBJ-memories.

Checkpoint text and marker controls also differ. A continuation request pulls chat history. Automatic latest-valid-marker selection then pushes a summary and selected tail; it is not another consumer request. The compaction model can see only image-stripped, budget-selected input, not necessarily the entire replaced prefix. Streaming persists a marker; generate retains only the in-process replacement through that channel. Both can feed later model-step invocations, but only the streamed path establishes stored-marker continuation. See MEM-RTE-chat-compaction and MEM-OBJ-compaction.

File-path reads and query-ID paging are pull. Their identifying arguments do not establish identifier-based push. Converted file display does not resolve opaque binary payloads or their producers. Saved evaluation reports have known file retention and structured/prose parts, but their later learning consumer is unresolved, not absent. See MEM-RTE-saved-files, EPI-RTE-query-results and the workflow annotation on RT-RTE-test-answer.

Curation and diagnostic memory

Consolidation follows the recorded transformation into reduced standing content or a continuation summary. Unproven faithfulness does not turn it into synthesis. Likewise, requested root-cause diagnosis does not establish a new claim absent from the inputs. Deterministic fingerprint collapse, existing finding updates and administrative withdrawal have their own wired curation warrants in EPI-RTE-finding-reconciliation. Explicit memory deletion is forgetting; supersession and disabling are withdrawal without deleting retained history. No automatic timed decay, semantic novelty or score-based promotion is inferred.

Recommendations are diagnostic memory even before any patch becomes live. The next diagnostic prompt receives compact prior summaries, not automatically all retained reasons. Reconciliation consumes fingerprint/status/count parts; dismissed matches supply an enforced veto and an admission criterion. This creates a qualifying trace-fed write/later diagnostic-consumer route without proving ordinary analytics improvement. Its mixed usage/tool-feedback input keeps partial source coverage. See MEM-OBJ-context-recommendation, RT-RTE-context-recommendation and EPI-RTE-finding-reconciliation.

PR creation, applied status and live context delivery remain different transitions. Applied can follow either reconciliation resolution or PR opening; it proves neither merge nor live consumption. Imported project revisions can supply instruction/knowledge only through the separate delivery and later prompt/read routes. Starting refresh chooses the source/workflow, not human approval of every admitted byte. See RT-RTE-context-delivery and RT-BAP-project-guidance.

Canonical resolution and coverage limits

Reconciliation supersedes EPI-CMP-memory-extractor with MEM-CMP-memory-extractor. That affects only the canonical citation for opaque extraction-provider state; it upgrades neither retention knowledge nor trace-learning strength. Other combined objects/routes remain valid alongside their parts. The current runtime and memory findings agree about budget-limited compaction input and the two applied status producers, so no old interpretation controls these classifications.

Every supported local positive survives in the units and derived union. Unresolved fork copies, dynamic skill/channel routes, story/app-state and file lifecycle, context providers, feedback filtering and provider memory are not removed to obtain complete coverage. No accepted bounded-absence record licenses a global trace-learning no or a trace-source inapplicability claim. Local pull units alone cannot make the full push-signal axis inapplicable.

Source-read log: none. This classification used only the supplied accepted members, their types, collection and record/worker contracts. No source file, incumbent profile or reference profile was read; no records or new evidence were added.