Claude Workstream Kit memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The scope is retained work memory, not every byte the runtime can access. File storage and intended repository retention coexist. The metadata retained for context-budget reporting is symbolic memory, not a dense model representation. Static installed instructions enter this profile only on the branch where experience changes standing guidance for a later consumer. External host memory and compaction are excluded responsibilities, not negative findings about the kit.
Symbolic form has separate consumption warrants. ACTIVE selectors, workstream index and telemetry have executable parser/selector consumers. PROJECT paths, detailed handoff metadata and archive locators instead have prescribed status/work/receiver or historical-retrieval paths, so their symbolic finding retains a claimed basis. Hook inbox count/age signals do not establish executable consumption of all handoff metadata. The known coverage assessment classifies these named forms; it does not upgrade prescribed consumers to wired ones. See MEM-OBJ-active-selector, MEM-OBJ-workstream-index, MEM-OBJ-context, MEM-OBJ-project, MEM-OBJ-handoffs and MEM-OBJ-archive.
The pointer's symbolic fields route selection; its prose directs later planning. Workstream claims supply context, while plans supply instructions and closure criteria supply validation conditions. Rewrite guards consume retained structure as criteria and enforce refusal only on their own replacement path. These forces must not be projected onto all model or shell writes. See MEM-OBJ-active-selector, MEM-OBJ-active-prose, EPI-OBJ-workstream-guidance and RT-RTE-rewrite-state.
Creation approval, substantive review approval and routine model admission are different write mechanisms. A generic CLI caller with a write flag does not establish a per-content human decision. Likewise a deterministic condensation of already-derived memory is not, by itself, another trace-learning route. The positive trace-learning findings concern experience-to-durable-checkpoint or governed-guidance pathways, with their claimed/afforded bases preserved. They establish neither parameter change nor demonstrated improvement. See RT-RTE-state-lifecycle and MEM-BAP-artifact-adoption.
Consolidation includes inspected selected-text reduction despite unproven semantic faithfulness. Rotation additionally reduces the default salience of spent content; historical retrieval remains requested, and full recovery depends on a faithful prior commit/tag. The synthesis request does not establish a new claim absent from its inputs. The EXTENDED protection applies to Learning condensation, not shared scoring or rotation. See MEM-RTE-learning-condense, MEM-RTE-state-rotate, MEM-CLM-marker-reopening and EPI-RTE-capture-synthesis.
Reconciliation supersedes the epistemic Learning-condensation declaration with MEM-RTE-learning-condense. All classification citations use that canonical declaration. The parent rewrite container remains valid; the supersession changes no value or evidence strength. The epistemic qualification-loss deduction concerns the same route, not an observed execution.
The automatic hook is one push chain with coarse eligibility/warnings and an identity match for active detail. Requested full-state and historical reads are pull; neither an identifier in a request nor a delivered response adds another push. The pointer is pushed whole, while full workstream prose is requested separately. See RT-RTE-session-reload and MEM-OBJ-archive.
Coverage remains partial where initial notes/PROJECT admission and provenance, destination-specific effects, exact transformation shapes, or qualifying mixed trace inputs are unresolved in the accepted records. A local positive does not erase those units. No absence record is supplied, so missing trace routes are not classified as no. Conversation origin supports session-logs; artifact evidence alone does not establish recorded tool-traces or trajectories. Raw evaluator logs afford later manual investigation, not automatic upstream learning. See MEM-OBJ-notes, MEM-OBJ-project, RT-RTE-acceptance-driver and EPI-RTE-capture-synthesis.
Source-read log: no frozen-source reads were needed. Classification used only the supplied accepted reports and reconciliation; no records, quotations or new evidence were added.