GBrain memory profile

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

Comparison rationale

This comparison classifies accepted records, not new source evidence. All positive findings are wired implementation witnesses; no supplied record establishes observed operation or causal benefit. Every axis remains partial because the accepted inventory retains included alternatives with unresolved parts, origins or consumers. An initial provenance gap does not erase a subsequent known transformation.

Checkpoint corpus and pointer manifest are separate units. MEM-OBJ-checkpoint-corpus retains rendered/redacted extraction input in files; MEM-OBJ-checkpoint-manifest retains SQL pointers and titles. The manifest is not a transcript summary or an opaque provider checkpoint. Harvested facts and later pointers supply durable context, so MEM-RTE-checkpoint-harvest supports trace learning without a claim of better recall. FIFO repair and inline best-effort publication remain different; banked text does not guarantee a surviving pointer.

Feedback is also split. MEM-OBJ-retrieval-events constrains rating eligibility, MEM-OBJ-retrieval-feedback controls idempotent update and retry accounting, and MEM-OBJ-ranking-weights changes later ordering. Event/history pruning can leave weights operative without complete attribution. The localized weights are symbolic numerical state, not dense-parametric model weights. MEM-BAP-ranking-weights supplies the downstream ranking witness, while the event/rating inputs have learning authority at the update consumer. This classification makes no improvement claim.

The accepted supersession maps EPI-RTE-fact-consolidation to MEM-RTE-facts-consolidation. All dependent findings cite the canonical memory record. Its strict reducer selects an existing claim and changes temporal eligibility: consolidation and invalidation, not novel synthesis. Checkpoint and feedback containers are preserved with independently declared parts, not superseded. Dream generation in EPI-RTE-dream-candidates supports a wired transcript-fed durable-page route but cannot establish the controlled curation value synthesize: output novelty or entailment is unestablished. Repair/quarantine instead supports invalidation through EPI-RTE-repair-retention.

Agency names admission, not the identity of whoever invoked the workflow. Software page/fact/skill admission is automatic where no per-content human approval is required. Initial authorship and owner-plane designation remain unresolved. EPI-RTE-owner-confirmation supplies a separate manual witness only on its interactive human-token branch; the remote scope branch is not evidence of human review. Neither branch proves truth.

Read-back classifies operations within chains. Requested evidence and polling are pull; unsolicited host additions and automatic internal selection are push. Core designation is coarse selection. Session/job/page-weight identity matches select retained material, while a slug printed in a pointer or requested filename alone would not establish identifier-based push. MEM-RTE-core-delivery, MEM-OBJ-checkpoint-manifest, RT-RTE-subagent-loop and MEM-RTE-feedback-update retain those distinct witnesses. Salient-entity detection does not specify enough selector mechanics to choose inferred-lexical, inferred-embedding or inferred-judgment.

Raw replay is not silently promoted to a learned strategy. RT-RTE-subagent-loop supports retained context/recovery; WF-RTE-agent-delivery supports a retained polled result, not semantic success. Published skill heads/files establish storage and revision, not installed instruction authority or executed assets. Opaque provider state and untraced skill optimization remain included uncertainties. No bounded absence records establish a universal trace-learning no or make trace_source inapplicable.

Trace-source correction separates checkpoint harvest from dream production. MEM-RTE-checkpoint-harvest and MEM-OBJ-checkpoint-corpus establish the checkpoint session-logs witness. EPI-RTE-dream-candidates establishes transcript-fed production without establishing agent-work-session origin, so its trace_source unit has no positive finding. Transcript fidelity requirements do not settle origin. This does not withdraw the separate dream trace-learning or trace-extracted lineage findings, whose definitions do not require agent-session origin. The manifest's readable titles are also classified as natural-language at later host consumption, separately from symbolic pointer selection.

Source-read log: none. This correction used the supplied contracts, accepted members, refusal and explicitly supplied previous profile as its baseline. It neither reopened source verification nor inspected any other incumbent or reference profile.