cass_memory_system memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile classifies accumulated memory and its later consumers, not all storage or permissions in the runtime. YAML/JSON/JSONL establish files; repository placement does not establish a separate repo substrate. Dense cached vectors establish parametric form, not model-weight learning or a vector database. The encoder's unresolved immutable identity remains a lineage limit even though the cache's derivation and similarity consumer are wired (MEM-OBJ-embedding-cache, RT-CMP-embedding-model). External cass is retained as an opaque history input rather than silently dropped to obtain complete coverage (RT-RTE-context-read).
Instruction and knowledge refer to different consumed parts. Rule imperatives supply standing guidance; descriptions, reasons and denial explanations supply context. Existing words, vectors and feedback also affect ranking or duplicate admissibility. Progress checkpoints route later work, while blocked content and trauma patterns can veto eligibility or matched actions. The latter's effective host prevention remains conditional and unobserved (RT-BAP-rule-guidance, MEM-OBJ-reasoning, MEM-OBJ-processed-log, MEM-OBJ-blocked-entry, RT-BAP-guard-denial). These forces do not authenticate source truth or prove useful behavior.
Reconciliation supersedes EPI-RTE-outcome-feedback with MEM-RTE-outcome-feedback; every classification of that route cites the canonical memory record. The remaining containers and parts retain their declared identities. Reconciliation's missing separate diary-status and validator-linguistic parts remain explicit inventory limits, not new record IDs or invented dossiers (EPI-OBJ-session-diary, EPI-OBJ-validation-result). The available combined records support the local status finding without resolving independently inventoried parts. Transient verdict/refinement objects are not classified as durable memory solely because they affect current admission.
Starting reflection, import or maintenance does not establish human content approval. Computational reflection and outcome grading are automatic; the trauma initializer's explicit batch approval is manual. Generic add/mark/replace/withdrawal callers remain unresolved at their actual decision points (RT-RTE-reflection-update, RT-RTE-onboard-add, MEM-RTE-portable-admission, MEM-RTE-trauma-admission). Human approval of scan results is not an automatic trace-learning write.
Curation follows the implemented effect. Duplicate checks can skip copies or combine feedback/provenance. Rule merge checks targets, creates a successor and retires sources, but its recorded structural effects do not establish a reduced representation of the inputs' existing claims; that branch has no positive curation finding and remains not-determinable (MEM-RTE-rule-curation). The independent fast-diary consolidation witness does not upgrade it. Replacement changes an entry; decay/maturity changes standing; retirement and healing withdraw reliance while preserving state. Inversion introduces an avoidance directive rather than merely shortening the original positive directive. Its synthesis is a policy transformation, not demonstrated acceptance of a new causal explanation (MEM-RTE-rule-curation, MEM-RTE-rule-withdrawal, EPI-RTE-curator-inversion). Warning-only conflict logging establishes no preventive contradiction gate. Model-requested extraction does not settle every opaque output's semantic transformation; wired persistence is not an upgrade of the requested consolidation into observed faithful consolidation.
Read-back is split by operation. Task context and diagnostic returns fulfill requests and are pull, even though code selects their contents. Reflection automatically supplies retained context to the delegated model; fixed update, checkpoint and safety consumers automatically receive relevant state. Those are push operations inside otherwise caller-started chains. Actual session/ID/hash matches establish identifier targeting; active-state availability is coarse; regex and blocked-token matching are lexical. Requested lexical/embedding context ranking supplies no inferred-embedding push witness (RT-RTE-context-read, RT-RTE-reflection-update, MEM-RTE-outcome-feedback, MEM-OBJ-blocked-entry, RT-RTE-trauma-guard). Exported host guidance remains afforded as instruction authority, but its later read-back direction has no positive finding: neither a host request nor automatic supply is recorded. This uncertainty leaves the wired requested export intact and also prevents classifying the later host's push signal (MEM-RTE-provenance-read).
Automatic summaries/rules and automatic standing updates qualify as trace learning because durable behavior-shaping results reach later consumers. They require neither new knowledge nor measured improvement. Original session turns and embedded tool material establish session-log and tool-trace source findings; a session-end callback is a trigger, not automatically the learned content. Asserted outcomes and delivery records do not establish linked trajectories or causal per-rule usefulness (MEM-OBJ-diary-entry, RT-RTE-reflection-update, MEM-RTE-outcome-feedback). No observed or causal basis is available, and no bounded absence record licenses a global negative.
Source-read log: no additional source reads. This classification uses only the supplied accepted records, their annotations and reconciliation; it adds no evidence, declares no records and reopens no report verification.