KHMS memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The strongest supported memory mechanism is automatic production of durable trace-derived proposals that later evaluation calls consume. Reviewed K-card adoption is a separate admission path. The retained intermediate candidates already meet the trace-learning write-route definition; observed improvement, parameter training and semantic truth acceptance are not required. MEM-OBJ-inbox and RT-RTE-mark-evaluation retain that chain. The weekly request for novel principles and own-mistake countermeasures stays at claimed semantic synthesis/learning basis: EPI-CLM-synthesis-review. Wired orchestration cannot upgrade a requested novel claim to a demonstrated transformation.
File-backed edges are not evidence of a graph database. NPZ numerical state is file storage with distributed-parametric form at its dense-ranking consumer; it is not a changed encoder model. MEM-OBJ-card-metadata and MEM-OBJ-embedding-index distinguish these. The initial corpus and exact encoder artifacts remain unavailable. Known derivation paths therefore do not establish complete lineage.
Authority depends on the consumer. Card facts provide knowledge; retained policy text provides instruction rather than an action veto. Trace inputs at proposal updates have learning authority; statuses, review coverage and session-state controls route or suppress code operations. Retraction text also supplies lint criteria, independently of its historical explanatory role. See RT-BAP-context-delivery, MEM-OBJ-mark-ledger, MEM-OBJ-hook-state and MEM-OBJ-tombstones. These forces do not prove model compliance or beneficial activation.
Explicit recall is pull. Event-driven card injection and code-assembled later evaluation packets are push, with targeted and coarse selectors kept separate. SessionStart supplies operational reminders, not a card search. Generated MEMORY navigation has an afforded loading channel; it contains neither card bodies nor the bootstrap-digest body claimed by older doctrine. See RT-RTE-recall-return, MEM-OBJ-hook-state and MEM-RTE-index-bootstrap. Tag matching on the automatic risky-action branch is an actual identity selector, unlike an ID merely printed on a requested file. MEM-RTE-precheck-return supplies that bounded witness.
No supersession affects these classifications: reconciliation establishes no duplicate replacement. The combined epistemic card and the memory body/metadata records remain valid at different granularities. The supplied doctrine/implementation distinctions remain local limits rather than competing whole-system verdicts. In particular, current-only search excludes archives, history returns loaded dead cards, and correction pointers are bounded. MEM-OBJ-card-metadata and RT-RTE-recall-return support those limits.
Missing initial-content control, private prompt contents, provider internals, deployed corpus and external version-retention evidence prevent complete coverage, not the positive classifications retained here. Generic reviewer identity does not establish manual content admission. No accepted bounded absence record warrants a local trace-learning "no"; unknown alternatives cannot be converted into negatives. The supported "yes" is an existence finding, not an exhaustive-value or improved-capacity conclusion. See MEM-OBJ-card-content, RT-RTE-operational-change and RT-CMP-agent-model.
Source-read log: no frozen-source reads were needed. Classification used only the supplied accepted members and their retained passages; no record, quotation or evidence was added.