ReasoningBank memory profile

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

Comparison rationale

This profile classifies accepted records only. No source files were read and no new evidence or records were added. The records' passages supply the warrants; no finding claims observed operation or causal benefit.

Inventory and uncertainty. The retained-object inventory separates MEM-OBJ-web-advice, MEM-OBJ-coding-advice, MEM-OBJ-lineage-ledger, MEM-OBJ-query-cache, MEM-OBJ-delivery-text, MEM-OBJ-trajectory and MEM-OBJ-feedback. Initial content control is not inferred from known subsequent writes. Exact vector producer/version lineage is unresolved, even though the inspected append computes new embeddings. Provider weights are separate computations, not accumulated memory updated by these routes. Ordinary episode histories and truncation without a durable summary are described by RT-RTE-context-control, but do not add a compaction learning route. Ancillary alternatives remain uncertainty units rather than disappearing from coverage.

Read-back is operation-relative. The memory member calls MEM-RTE-web-retrieval and MEM-RTE-coding-retrieval host-triggered requested setup rather than unsolicited background push. That source-native characterization is preserved. Under this profile's definition, the runner pulls from its selector while automatic setup pushes selected advice into a task model that did not request retained material. The initiating task request is not itself a memory request by that consumer. The same chain therefore supports both directions without counting a requested selector return as an extra push. Similarity ranks cached queries; a subsequent task-ID equality really selects the delivered row. Both selector operations matter. The bank's status and template labels do not establish quality filtering or deduplication.

Curation and learning are separate. The reports distinguish extracting from traces from consolidating or revising old advice. The profile's broader consolidation definition includes reducing retained traces, so ordinary RT-RTE-web-admission and RT-RTE-swe-admission support that operation. The absence of generated propositions prevents deciding whether any additional claim is synthesized; instructions to generalize do not settle that question. Automatic trace-fed durable advice plus later delivery establishes architectural trace learning, independently of improved capacity, actual decision dependence or a parameter update. Retained experience has learning authority at the updater; advice has instruction authority downstream, dense vectors have ranking force, and lookup/outcome fields route specific operations. These are separate consumers, not interchangeable labels.

Alternative routes retain their own evidence strength. RT-RTE-web-scaling affords intended retention but has repeated-directory inputs, client invocation and output-shape defects. Its requested prose cannot classify an entire incompatible tuple payload. MEM-RTE-standalone-induction has an append body but unresolved imports/launch context and no established later destination consumer. Neither inherits the ordinary loops' wired trace-learning finding. Intended trajectory inputs can be classified at affordance strength while the complete learning chain remains unresolved.

Canonical identities. Reconciliation supersedes EPI-RTE-query-selection with MEM-RTE-web-retrieval, and EPI-RTE-legacy-induction with MEM-RTE-standalone-induction. This profile uses those canonical routes; their findings remain bounded to wired ordinary selection and afforded standalone induction respectively. The epistemic web/coding memory containers remain distinct from their content and ledger parts, so their semantic uncertainty can be cited without treating them as duplicate routes. No supersession upgrades a local basis or removes uncertainty.