Basic Memory memory profile

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

Comparison rationale

The profile classifies accepted source records, not a deployed configuration or an observed run. Every axis remains partial because the accepted inventory retains material alternatives whose control, storage, consumption or transformation is unresolved. A wired witness establishes its local path, not complete coverage. Provider inference components support the numerical embedding distinction, but do not establish accumulated provider-side memory or parameter learning.

The schema-report duplicate is resolved to EPI-RTE-schema-validation. This is the only reconciliation supersession affecting the classifications; the transient validation result is not itself a retained certificate. Stored schema criteria can supply validation authority without a write veto or factual endorsement.

The accepted RT-RTE-pi-checkpoint, RT-RTE-hook-checkpoint, MEM-RTE-tau-handoff and MEM-RTE-hermes-continuity connect automatic trace-fed writes to later context consumption. They therefore support trace_learning yes even though the reports establish neither improved capacity nor successful self-improvement. Tau uses a model and prior handoff; Pi/Hermes use fixed extraction. Both can consolidate existing claims. The word synthesis does not establish creation of a claim absent from inputs. EPI-RTE-reflect-notes supports separate claimed consolidation/evolution/invalidation and a claimed trace-fed continuation, not wired execution.

Raw MEM-RTE-trace-archive retains and rereads operational metadata, prunes by age, and counts replay keys. It neither merges duplicate knowledge nor establishes a promotion pipeline. MEM-RTE-note-withdrawal deletes accepted state; only the separately prescribed archive/status path supports history-preserving invalidation. These distinctions prevent unsupported dedup or promote values.

The direction and signal axes split internal requested reads from automatic host supply. MEM-RTE-openclaw-context provides coarse task/recent-activity recall through systemPromptAddition; that channel alone does not turn retained reference content into instruction. Tau branch receipts actually select the pushed checkpoint; a requested note path alone would not establish identifier push. Pi prompt-driven search is recorded, but its query alone does not resolve lexical, embedding or model-judgment selection; that uncertainty remains beside the known coarse eligibility filters.

Retained reasons are local: Tau requests decisions and rationale and passes earlier handoffs to later synthesis; presence and faithfulness remain unobserved. Generic Markdown can carry reasons without requiring them. QA rejudging is a later consumer of stored cases, but the separately described benchmark file writer cannot be joined into a particular QA persistence or product-update loop. See MEM-RTE-tau-handoff, MEM-OBJ-markdown-content and EPI-RTE-qa-judge.

The storage classification of MEM-OBJ-graph-publication covers published graph rows consumed by RT-RTE-context-build, not the process-local prepared payload in the current mutation. Its retention into a later memory invocation is not established. That transient payload is excluded across the profile's classifications; retained graph and publication metadata remain included. The independent cross-invocation OpenClaw recall cache still supports in-memory storage through MEM-RTE-openclaw-context.

Source-read log: no source checkout files were read. Classification used only the supplied accepted members and reconciliation. No records, quotations or execution evidence were added.