ByteRover CLI memory profile

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

Comparison rationale

This profile classifies the accepted records, not a fresh source inspection. All positive bases are implementation wiring local to the named part or mechanism; none claims observed activation, factual acceptance or improved performance. The inventories are partial because the records retain alternative writers, compression branches and consumers without establishing a complete account of their retained objects and routes.

Retained content, access state and authority

Topic prose, typed selection markup and runtime signals have different consumers. Rule/decision text can supply directions to later agents; statements and reasons supply knowledge. Neither becomes enforcement merely because the renderer preserves a severity label. Search treats retained text and access metadata as ranking inputs. Numerical sidecar updates also form a trace-fed behavior-shaping route. Their tier label does not validate the underlying claim. See MEM-OBJ-bv-rule, MEM-OBJ-bv-reason, MEM-RTE-search-recall and MEM-RTE-signal-feedback.

The same sidecar has a KV interface and file backing; these are layered substrate witnesses, not two independent memories. Compaction's logical message key store has no stronger physical-backing classification here. Model execution does not establish retained model-weight updates. See MEM-OBJ-runtime-signals, MEM-OBJ-compaction-boundary and RT-CMP-content-generator.

Review statuses have enforcing force at legacy Cloud export, while recovery metadata routes rollback. The static writer schema is not accumulated memory and therefore does not create a validation-authority finding in this profile. Pending review does not prevent local retrieval. See EPI-RTE-review-decision, RT-RTE-cloud-sync and EPI-RTE-dream-undo.

Admission and transformations

Software admission is automatic where no explicit per-content human decision is required. This does not deny possible human authorship. Generic callers, editor clients and reviewers do not establish manual control; those paths retain unresolved principal identity. Installation or workflow initiation is not approval of every retained part. See RT-RTE-curate-session, RT-RTE-agent-curation, RT-RTE-direct-edit and RT-RTE-hub-install.

Consolidation covers reductions of existing retained content and activity, including deterministic checkpoints. It does not require novel knowledge or faithful output. Semantic merge/synthesis requests do not establish dedup or synthesize operations. Archive withdrawal preserves former bytes and is distinct from irreversible deletion; imperfect HTML-cache invalidation prevents claiming immediate removal from every consumer. See MEM-OBJ-summary-index, MEM-OBJ-checkpoint-progress, RT-RTE-dream-maintenance, MEM-OBJ-dream-archive and MEM-OBJ-query-cache.

Request, automatic supply and continuation

Requested query, search, focused read and expansion are pull operations. Ranking a requested return does not make it push. Internal manifest context is a separate automatic supply operation within a requested execution: code selects broad entries without a request for those particular retained parts, and the resolver ignores query relevance. Its coarse lane/budget selector is not search's activity ranking because this constructor omits the signal store. See RT-RTE-query-retrieval, MEM-RTE-topic-read, MEM-RTE-archive-expansion and MEM-RTE-manifest-injection.

The full-summary product has a local automatic loading witness in MEM-OBJ-compaction-boundary. Reconciliation leaves a wording conflict with RT-RTE-context-control: the runtime says only post-boundary messages reload, whereas the memory record includes the summary boundary itself. This profile does not settle that conflict. It retains the local memory-record positive and marks the affected reload units partial rather than asserting one complete cross-report history inventory. Rolling checkpoint JSON is separate: the continuation asks the model to read it, so payload retrieval is pull, not automatic dereferencing. See MEM-OBJ-checkpoint-progress.

Signal feedback, full compaction and rolling checkpoints qualify independently as trace learning: activity feeds automatic writes, retained results shape later ranking or task context, and within-task persistence can suffice. This says nothing about observed improvement. Search hit activity and checkpoint tool calls support tool-traces; conversational message content supports session-logs. Unenumerated mixed inputs remain unresolved. No event-stream or trajectory source follows merely from callbacks, timestamps or linked message identities.

Canonical records and coverage limits

The supplied reconciliation supersedes the epistemic runtime-signal declaration with MEM-OBJ-runtime-signals, and the epistemic focused-read declaration with MEM-RTE-topic-read. All classifications use those canonical destinations. Supersession changes identity resolution, not value or strength. The topic container, its semantic/selection parts, the context-control route and its two retained products remain distinct.

Markdown-oriented summaries/manifests and HTML-oriented dream maintenance cannot stand for one universally covered memory plane. The archive cue/full service is not the dream archive, and exposed expansion is not an observed automatic archive lifecycle. Registered ingest/map/swarm paths, alternate compression, version/worktree/migration replacements, downstream log consumers and external host placement prevent stronger complete-inventory claims. These are unresolved included alternatives, not absence findings. See MEM-OBJ-summary-index, MEM-OBJ-archive-stub, MEM-OBJ-full-archive, RT-RTE-agent-curation, RT-RTE-cloud-sync and RT-RTE-context-control.

Source-read log: no target source files read. Only supplied invocation, governing contracts and pinned members were read; no new evidence, quotations or analyst records were added.