Spacebot memory profile

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

Comparison rationale

This is a classification of accepted records, not a new source inspection. All positive findings above are wired capacities. No supplied record establishes observed operation or causal benefit. A partial axis preserves those capacities while keeping unresolved included parts visible. Whole-system functional scope does not turn selected source inspection into complete memory coverage.

Canonical references follow reconciliation: the memory-declared encoder, claim content, selectors and wiki page replace the duplicate epistemic component/object declarations. Their paths are MEM-CMP-embedding-model, MEM-OBJ-memory, MEM-OBJ-memory-metadata and MEM-OBJ-wiki-page. No supersession changes a value or upgrades a basis. Chronicle retention remains a part of context control, and completion admission remains a part of delegated work; neither container is discarded.

Dense embeddings have parametric form and vector substrate. SQL association rows have symbolic form, graph organization and SQLite persistence. Neither new vectors nor changed claim text establish model-parameter training. Exact encoder provenance remains unresolved in MEM-CMP-embedding-model. External provider internals are outside this profile, not classified as absent memory.

Trace-fed summaries qualify even when they compress existing claims rather than add knowledge. The operative chain is retained experience, automatic derived write, then a later consumer. This holds for conversation extraction, temporal narratives, chronicles and procedure reflection; session-live rolling summaries have a shorter horizon than durable checkpoints. These findings do not inherit the stronger meaning of successful learning or self-improvement. See MEM-RTE-temporal-synthesis, MEM-RTE-chronicle-retention and RT-RTE-skill-change. Fallback notices and unsummarized emergency cuts are not silently treated as successful summaries.

Learning authority belongs to experience at an update consumer; knowledge or instruction belongs to the resulting artifact at its later work consumer. Ranking belongs to retained text, vectors and metadata used by selectors. Forgotten flags and local pin/task/terminal checks have enforcing force only at their named consumers. They establish neither truth acceptance nor global protection. See MEM-BAP-memory-selectors, RT-BAP-skill-instruction and WF-RTE-completion-admission.

Consolidation and deduplication describe implemented selection/compression/merge mechanisms, not universal semantic fidelity. Substring loss, first-line omission or a model extraction error does not establish synthesis. The maximum-importance assignment in EPI-RTE-batch-consolidation supports salience promotion of the survivor; it does not establish epistemic endorsement. Novel claims in particular model outputs remain unresolved, as EPI-RTE-memory-extraction explains.

Requested recall uses lexical/vector/graph algorithms, but its fulfillment remains pull. Automatic participant matching and required-skill name resolution establish identifier push because they select delivered retained parts. Availability, chronological windows and budgets separately establish coarse push. A stored label alone establishes neither direction nor signal. See MEM-RTE-memory-recall, MEM-RTE-context-supply and RT-CLM-required-skills.

Initial authored content and generic administrative identity do not settle write agency. No manual value or negative trace-learning finding is manufactured from those gaps. Full attachment payloads, org seeding/refresh, auxiliary state revisions and restored-worker alternatives remain unresolved. Known displays and local routes do not license complete-value statistics for those surfaces.

Additional source reads: none. Classification used only the supplied accepted members and their records; no quotes or new records were added.