MemoryOS memory profile

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

Comparison rationale

All findings classify accepted records at the frozen revision. The strongest basis used is wired, not observed operation or causal benefit. No source files were read for this profile; no new records or evidence are added.

The comparison keeps original dialogue, derived text, dense selection material and symbolic maintenance state separate. File encoding does not establish a vector substrate; the FAISS index and ChromaDB collections do. External model weights are dependencies, not accumulated memory changed by these routes. Initial-file and direct-caller provenance does not establish human approval. The known automatic derivations therefore coexist with unresolved origins (MEM-OBJ-memory-access, MEM-RTE-direct-update).

The reconciliation supersedes the epistemic profile, user-knowledge and assistant-knowledge objects with their memory counterparts. This profile uses MEM-OBJ-user-profiles, MEM-OBJ-user-knowledge and MEM-OBJ-assistant-knowledge throughout. Raw dialogue and summary parts remain valid alongside the larger memory containers; they are not treated as additional independent systems.

Learning authority at the update consumer is distinct from knowledge or instruction at the answering consumer. Dialogue and old profiles feed durable updates; resulting preferences and assistant traits then guide later answers. Retained embeddings/text influence ranking even though scoring code is fixed. Analyzed flags determine extraction eligibility, supplying a validation criterion and an enforced exclusion, not evidence of factual acceptance (RT-RTE-profile-analysis, MEM-OBJ-memory-access, RT-BAP-user-context, RT-BAP-assistant-context).

Extraction of standing facts is consolidation. Unchecked model interpretation does not establish synthesis: the accepted epistemic account leaves semantic relations indeterminate. Profile replacement is evolution, and capacity/LFU removal is decay, not history-retaining invalidation. Tier migration and heat-selected long-term admission establish promotion without establishing truth or improvement (MEM-OBJ-mid-term, MEM-OBJ-user-profiles, RT-RTE-profile-analysis, EPI-OBJ-summaries).

The ChromaDB correction matters locally. New-session insertion overwrites the JSON page backup before saving; original vector text can still be recalled, but those initial pages cannot feed backup-only extraction or chain overview enrichment. Later merges can retain new backup pages. Their retained analyzed flags are checked before extraction: the flag is an eligibility criterion (validation), and a marked page is excluded (enforcement), as well as selecting the input window (routing). These control effects have their own known unit, separate from text at response consumers, ranking inputs and dialogue at update consumers. Positive trace learning therefore rests on dialogue-derived summaries and eligible surviving-backup updates, never an assumed intact initial-page extraction chain. Dependency collection-add behavior remains unresolved separately (MEM-OBJ-chroma-storage, RT-RTE-chroma-dialogue).

Automatic selection within a requested response is still pull. MCP/browser requests and benchmark/scorer reads are also pull; no extra push is counted for formatting their returns. The remaining uninspected driver/helper coverage prevents a global pull-only assertion and leaves push-signal coverage uninspected rather than globally inapplicable (RT-RTE-dialogue-response, RT-RTE-mcp-access, MEM-RTE-profile-display, RT-RTE-locomo-evaluation).

Trace learning means a qualifying automatic conversational write and later consumer, not demonstrated better capacity or weight training. Imported benchmark conversations remain session-log sources. The answer/reference ledger has a scoring consumer without an established feedback update. The accepted records supply no canonical bounded-absence warrant for a controlled local no; its trace-learning and trace-source units remain not-determinable. Unknown upstream/direct producers cannot inherit the known conversational source (MEM-OBJ-evaluation-memory, MEM-OBJ-loco-results, RT-RTE-locomo-evaluation).