MemWiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The operative memory is the mutable project wiki and its original-input directory, not the entire installed prompt surface. The current and legacy CLIs seed authored starter files, but the consequential cross-session writes and reads are requests to an external foreground agent. Hence wired seeding does not upgrade claimed model admission, condensation or consumption. The inspected retained prose establishes available representations, not observed production or beneficial activation. See RT-RTE-cli-init, RT-RTE-legacy-init and RT-RTE-agent-session.
The index supplies routing through model-interpreted page labels and links, not a symbolic resolver or retrieval ranker. Hot next steps and coding conventions supply instruction; facts and architectural rationale supply knowledge. The work log has a separate update consumer in fold: retained experience feeds another durable context artifact, while its later reader receives contextual knowledge. None of these paths establishes permission enforcement or epistemic acceptance. See MEM-OBJ-index, MEM-OBJ-hot, MEM-OBJ-domain-pages, MEM-OBJ-decisions, MEM-RTE-log-fold and EPI-ABS-acceptance-gate.
Work-derived hot/log writes and fold satisfy the trace-learning axis at claimed route level: automatic content admission, durable derived context, and an identified later-consumer role. This classification is independent of improved capacity, source truth or successful execution. It does not license a trace-source guess: the recorded chain starts from coding experience or already-summarized work, not evidenced conversational turns, tool results or event callbacks. Repository-reference acquisition is outside qualifying trace-fed writes; this does not establish a bounded trace-learning absence. Raw inputs may contain activity or static reference, but no actual payload resolves that alternative. See RT-RTE-agent-session, MEM-OBJ-log, MEM-RTE-log-fold, MEM-RTE-repository-ingest and MEM-OBJ-raw.
Consolidation follows the requested work-summary and older-history reduction. A named synthesis command does not establish novel claims. The epistemic ledger leaves the exact semantic relation indeterminate without paired inputs/outputs; therefore the raw/repository branch cannot establish controlled synthesize, while fidelity uncertainty does not erase the explicitly requested fold consolidation. Lint may end at proposals, so its correction admission and curation stay unresolved. See MEM-RTE-log-fold, MEM-RTE-wiki-lint and MEM-RTE-repository-ingest.
Read-back remains pull within the scoped shipped routes. The foreground agent requests hot and conditional index/domain files; fold requests retained history. Hook discovery requests those reads but does not independently select and inject retained content. The push-only signal axis is therefore inapplicable here, not unknown because of the excluded host's private context machinery. See RT-RTE-agent-session, MEM-OBJ-raw, MEM-RTE-log-fold and MEM-RTE-wiki-lint.
Reconciliation supersedes the epistemic hot/index/log and lint/fold duplicates with their memory declarations; this profile cites those canonical memory records. The epistemic four-domain-page container remains valid but overlaps the memory domain/decision parts rather than replacing them. These identity decisions change no comparison value. The retained unresolved index-count conflict prevents a consistent numerical link inventory, not filesystem/form/routing or direction classification: no finding here depends on that count. The shared missing landscape/competitor links remain a limitation of reported delivery, not included opaque retained objects. See MEM-OBJ-index, MEM-OBJ-log, MEM-OBJ-domain-pages, MEM-OBJ-decisions and RT-RTE-agent-commands.
No source files were read for this profile. Classifications use only the supplied accepted records and their annotations, resolved through supplied reconciliation. Missing original payloads, exact sentence provenance, trace content and lint-admission facts prevent stronger coverage assessments. External host behavior is an explicit boundary exclusion, not a silently discarded internal alternative.