OpenRSI memory profile

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

Comparison rationale

The profile classifies accepted records, not fresh source evidence. All positive findings distinguish wired local mechanisms from the external RL training consumer, whose parameter-update use is afforded by the integration rather than an inspected optimizer. There are no observed or causally supported findings. Every axis remains partial because its included inventory retains unresolved alternatives; a positive witness does not establish complete enumeration.

The journal, cards, strategy board and rich-summary cache are different retained parts. The source account records distinct persistence and recovery paths in MEM-OBJ-journal, MEM-RTE-experience-cycle and MEM-OBJ-rich-summary. Separate cache alternatives are included without assuming content freshness or complete cross-file recovery. The reconciliation supersedes EPI-OBJ-rich-summary with MEM-OBJ-rich-summary; all object findings therefore use the latter canonical record. The epistemic production/retention/consumption routes remain independently usable. The combined EPI-OBJ-experience-card container was not superseded; this profile uses the separate card and board records where their parts differ.

Knowledge at an operator prompt is different from learning authority at an update consumer and executable source at a task executor. Parent program text can supply examples while metrics rank it and status flags control eligibility. These are distinct effects, not a blanket instruction or enforcement label on all stored experience. See MEM-BAP-experience-guidance, RT-BAP-training-parameters and RT-BAP-candidate-execution.

Trace learning here denotes an automatic trace-fed retained update with a later consumer, not improved capacity or accepted explanatory knowledge. The Evo summary/card chain and RL retained-context chain qualify independently of whether trained parameters are ever deployed to Evo. Model-summary novelty remains unresolved: neither requested reusable experience nor failure to prove entailment establishes synthesis. Deterministic reduction and truncation support consolidation locally, without certifying faithful retention of qualifications. See EPI-RTE-summary-produce and EPI-RTE-summary-consume.

The SFT launcher admits arbitrary caller-selected data. Its known rollout-derived branch does not settle that alternative's historical lineage, trace origin or human admission control. Likewise no manual finding follows just from editable files, caller identity or a launch decision. Pruning is forgetting from the inspected database, not historical invalidation of every copy. See RT-RTE-sft-training, MEM-RTE-sft-reuse and MEM-RTE-rl-reuse.

Pull describes requested restoration or sample reads; push describes the host's selection and unsolicited supply of old work to the next proposal model. These operations can coexist in one chain. Identifier findings refer to actual parent/task matches selecting delivered content, not to a cache filename alone. Coarse findings refer to tails, broad journal slices and fixed score/status/cap filters, not inferred semantic similarity. Complete debug/crossover payloads, RL helper details and untraced alternatives prevent a stronger direction or signal inventory. See MEM-RTE-operator-context.

Quality reports are retained outputs, but their immediate recommendation/CLI consumers do not establish a later accumulated-memory route. Their later consumption remains unresolved rather than being assigned knowledge or enforcement authority by analogy. See EPI-RTE-quality-recommendation. Static task skills are excluded from the comparison's accumulated-memory scope; the inspected injector does not establish their trace-fed maintenance. See MEM-OBJ-task-skills.

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