AI-Context-OS memory profile

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

Comparison rationale

The profile classifies accepted records, not the runtime's entire storage surface. Canonical prose and symbolic metadata share files but have different consumers. Usage JSON is a separate durable ranking input; the SQLite request history instead feeds diagnostics and maintenance advice. Graph-based scoring does not turn either file storage into a graph substrate. Current chat turns and transient query results are not additional accumulated memory. See RT-OBJ-memory, MEM-OBJ-memory-meta, MEM-OBJ-memory-usage, MEM-OBJ-observability-db and RT-RTE-desktop-chat.

Evidence bases and coverage

All implemented findings are wired, not observed or causally supported. External adapter discovery and instruction consumption have an afforded host-reader path; a generated file does not prove a host loaded it. The same value can therefore retain separate wired and afforded findings. The records do not establish accumulated provider state behind the inference interface; provider inference's parametric form is not evidence of a parametric memory write. Opaque attachments retain unknown representation and consumer force rather than inheriting the readable sidecar's classification. See RT-CMP-inference-provider, MEM-OBJ-inbox-attachment, MEM-OBJ-router-manifest and RT-RTE-adapter-handoff.

Known local scopes do not establish complete system coverage. Initial vault provenance, generic caller control, incomplete alternate mutations, session/scratch consumers and possible provider-side accumulation remain unresolved where they prevent a conclusion. They do not weaken a distinct known implementation witness. Storage, form, lineage, authority, agency and transformation coverage are assessed independently.

Admission, curation and authority

Human confirmation establishes manual governance deletion/boost, not manual control of every callable mutation API. Conversely, invoking a save or analysis request does not make the code's subsequent counters, projections, candidates or derived tag writes manual. A generic apply caller does not establish a human per-content decision. These distinctions follow MEM-RTE-governance-actions, EPI-RTE-importance-boost, MEM-RTE-journal-derive, MEM-RTE-onboarding-write and RT-RTE-ingest-proposal.

Curation follows actual transformations. Normalization preview reduction from retained item text and extracted task lines support consolidation; explicit content replacement supports evolution; importance boost supports promotion. Acquisition and discovery/index rebuilding alone establish no curation operation. Deletion forgets rather than retaining an invalidated historical memory. Proposal rejection has narrower history-preserving withdrawal from pending/applied operational subsets, but apply does not enforce an irreversible veto. Duplicate-discard recommendations and exact-title creation suppression do not establish merging/removal of retained duplicates. Model proposal semantics remain indeterminate rather than automatically synthesized. Optimization apply only changes a status, not the proposed compression, merge or priority edit. See MEM-RTE-inbox-capture, MEM-RTE-journal-derive, EPI-OBJ-proposal-content, EPI-OBJ-proposal-control and RT-RTE-optimization-status.

Authority is consumer-specific. Lexically scored L0 has a symbolic feature-consumption path as well as natural-language interpretation. Protection supplies validation and enforcement at guarded mutation consumers, not universal filesystem control. Journal tag writes and restore are alternate admissions outside that guard. Retained usage events have learning authority at the updater; derived counts have ranking authority downstream. Persisted advice has knowledge authority at maintenance readers, not executable authority to carry out its suggested change. See MEM-OBJ-l0, RT-RTE-workspace-write, MEM-RTE-backup-restore, MEM-RTE-access-feedback and RT-RTE-optimization-status.

Read-back and trace-fed updates

Requested query selection and ID-directed retrieval are pull, even when fixed code ranks or expands dependencies. Requested returns are not reclassified as push merely because software chooses or delivers their contents. The accepted records do not establish a separate unsolicited selector for these routes. Signal is therefore locally inapplicable on assessed pull chains, but not globally inapplicable: missing session/scratch/binary consumers and opaque provider behavior leave alternative directions unresolved. Catalog IDs, requested skill IDs and lexical query scores alone do not establish identifier or lexical push. See RT-RTE-context-query, RT-RTE-skill-load, MEM-OBJ-session-log, MEM-OBJ-inbox-attachment and RT-RTE-adapter-handoff.

Trace learning has a supported local yes for automatic exposure-to-usage-to-later-ranking and for exposure-to-durable-maintenance-advice-to-maintenance-reader. The latter does not imply automatic canonical repair. Their original inputs are discrete read/serve/request events, not recorded conversational turns, complete tool traces or multi-step trajectories merely because MCP participates. Both retain wired rather than observed evidence. No new knowledge or measured benefit is required by this axis. Journal derivatives and arbitrary proposal inputs remain unresolved as trace-fed routes because their original content is not uniformly established as experience; the separately retained session summaries lack a later reader. No evidenced-absence record supports a local no, and no system-wide no is derived. See MEM-RTE-access-feedback, MEM-OBJ-observability-db, RT-RTE-optimization-status, MEM-RTE-journal-derive and MEM-OBJ-session-log.

Canonicalization and source-read log

Reconciliation resolves the duplicate usage-store, usage-feedback, capture-route and metadata records to MEM-OBJ-memory-usage, MEM-RTE-access-feedback, MEM-RTE-inbox-capture and MEM-OBJ-memory-meta. These canonical destinations are used throughout. Supersession changes citation identity, not the wiring basis, coverage, source-warrant limits or distinctions between access and success. Canonical containers, their material parts and the intersecting epistemic content/instruction records remain distinct; no stronger union is inferred from overlap.

No target-source file was read for this profile. Classification uses only the supplied accepted members and reconciliation. No new record, quotation, source ID or execution evidence is introduced.