Incremental self-improvement memory profile

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

Comparison rationale

This profile classifies the documented memory boundary, not a verified executable implementation. All positive mechanisms use claimed: the accepted records supply design and attributed operation, without inspected code or independently observed traces (RT-RTE-instruction-cycle, RT-RTE-credit-assignment). Coverage and evidence strength are separate. A fully described array can have known substrate at claimed strength; unresolved navigation representation and arbitrary work-history paths keep the system axes partial. No missing execution evidence is converted into absence.

The enduring learning product is the live probability table, not a successful program archive. Its localized entries have fixed sampling/update rules, so they are symbolic rather than distributed-parametric. The stack splits saved policy vectors from outcome and access metadata. The latter feeds comparisons and selects recovery records; it does not retain an explanation or complete trajectory (RT-CMP-probability-distributions, RT-OBJ-stack). Initial fixed numerical construction is other-compiled lineage; supplied observations are imported; retained event measurements are trace-extracted. These distinctions do not recover the semantic origin of every arbitrary integer.

Pi has relative weighting and operation-selection effects at the sampler. GetP instead supplies reference data. Stack evidence has learning authority at the automatic policy-retention consumer, while restored Pi changes downstream selection. The variable scorer supplies outcome feedback rather than approving individual writes. These paths do not establish explanatory knowledge acceptance, reliable action or causal benefit (RT-BAP-probability-selection, MEM-RTE-getp-read, EPI-RTE-rate-comparison, WF-RTE-variable-cycle).

Modifier writes evolve policy. Rejection restores distributions and pops history; because the records do not establish preserved invalidation history, this removal is decay rather than invalidate. Periodic external-array zeroing and proposed circular history overwrite also forget scoped retained content. Copying snapshots establishes neither reduced-claim consolidation nor novel-claim synthesis (RT-RTE-self-modification, RT-RTE-credit-assignment, RT-OBJ-stack, MEM-OBJ-variable-array).

Requested operand/GetP/Read retrieval is pull. Automatic sampler policy supply, stack evaluation and periodic scoring are push operations with different selectors. Address identities select actual delivered distributions or grouped recovery records; an address in a pull request is not push evidence. The scorer reads the whole array on a timer, giving a coarse selection rather than targeted relevance retrieval (RT-RTE-instruction-cycle, MEM-RTE-variable-readback, RT-RTE-credit-assignment, WF-RTE-variable-cycle).

The automatic historical reward/time keep-or-restore route qualifies as trace learning: numerical experience feeds durable behavior-shaping state used by later sampling. Its source is discrete payoff and runtime events, not a required retained transcript. This mapping does not assert improved capacity or upgrade the reported comparisons to causal support. Other modifier inputs and arbitrary event encodings remain unresolved; no local no is invented to complete that inventory (RT-RTE-credit-assignment, RT-RTE-self-modification, RT-OBJ-work-area).

Reconciliation supersedes EPI-RTE-getp-read with MEM-RTE-getp-read. The profile uses that canonical route while retaining its distinct output object EPI-OBJ-getp-values. This affects policy snapshot lineage, authority, admission and pull classification, not their evidence strength. The work-area container and its input/snapshot parts remain valid; neither those parts nor EPI-RTE-rate-comparison supersede their containers. No other supersession affects a value.

Proposed instruction reuse, teacher protection and circular overwrite retain only their declared force. Human teacher protection is a separate proposed retention decision, not a manual base write gate. Arbitrary work-cell semantics, the navigation environment's retained representation, exact external Read destination, and exercised GetP or Write-to-Read use remain limited by the accepted records. Those missing facts prevent complete lineage/consumer inventories and stronger operation claims; they do not erase supported mechanisms (RT-OBJ-program-area, RT-CLM-usefulness-invariant, MEM-RTE-variable-readback).

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a direct capture read.