ClawVault memory profile

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

Comparison rationale

The profile classifies accepted records, not a new source inspection. All positive findings are wired unless a finding explicitly identifies an afforded consumer. No record supplies observed operation or causal benefit. Axis coverage is partial even where several units are known: the memory report inventories included alternative mutations and consumers without complete part-level coverage. Those alternatives remain named rather than being removed to obtain complete statistics.

Objects, forms and origin

RT-OBJ-vault-documents is a container for MEM-OBJ-observations and EPI-OBJ-entity-profile, not another copy of either family. Prose, fixed-rule metadata and dense vectors retain separate form findings. Vector cache serialization remains unresolved in MEM-OBJ-embedding-cache; vector existence does not establish SQLite or provider-object storage. Provider weights are inference dependencies, not established use-derived retained memory. Raw strings and the separate graph snapshot remain in the inventory without inventing later foreground readers.

Reconciliation supersedes EPI-OBJ-extracted-facts with MEM-OBJ-fact-store, and EPI-RTE-replay-import with MEM-RTE-history-replay. The profile cites the canonical targets. These identity amendments do not erase the distinct epistemic extraction, supersession and context-selection findings. EPI-RTE-fact-extraction, EPI-RTE-fact-supersession and EPI-RTE-fact-context refine rather than supersede the combined memory route.

Force and admission

Existing observations condition later compression, giving them learning force at that update consumer; their later snippets separately supply knowledge. Fixed-code scoring of retained prose/vectors establishes ranking without a retained learned ranking policy. Persisted processed hashes supply an eligibility criterion and veto at maintenance admission; fact validity controls active delivery. These local code effects do not make delivered prose an enforced truth. See RT-BAP-worker-state, RT-RTE-observe-session and EPI-RTE-fact-context.

The optional interactive handoff permits a human to supply retained content, unlike generic direct store, patch or checkpoint callers whose control remains unresolved. The later Git approval does not retroactively approve the memory write. RT-RTE-session-handoff, with its workflow annotation, retains this distinction. Automatic writes elsewhere do not resolve these caller-dependent mechanisms.

Curation follows transformations, not names. Compression/extraction consolidates even where fidelity fails; the concrete kind conjecture in EPI-RTE-entity-synthesis independently supports synthesis. RT-RTE-reflect-week supports weekly consolidation/promotion, not theory criticism by its name. MEM-RTE-observation-archive establishes only local age-based withdrawal; derivatives and other enumerators prevent universal forgetting.

Delivery and trace updates

Requested wake, tools, explicit context and inject remain pull even when code selects their returned parts. The unsolicited prompt hook and outbound lookup are push. Prompt-derived relevance, optional compatible vectors, recency/availability caps and session-key recap lookup have separate signal warrants in RT-RTE-prompt-recall. Path labels on selected snippets are not another identifier selector. Optional LLM matching on a requested inject path is not evidence of judgment-based push.

Automatic observation, replay, capture, fact extraction and weekly consolidation provide qualifying trace-to-retained-context chains. This is the profile's trace-learning definition, not demonstrated capacity improvement. Raw capture alone supplies no such conclusion. Original conversational content supports session-logs; neither a hook callback label nor normalized journal-like serialization proves environmental event origin. Mixed hook/tool payloads, normalizers and alternative writers remain unresolved under MEM-OBJ-raw-transcripts, MEM-RTE-history-replay and EPI-RTE-fact-extraction.

Source-read log

No frozen source files were read for this classification. Only the supplied contracts and pinned members were read; no ambiguity required an authorized source reread. The profile adds no records, quotations or evidence.