Binder memory profile

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

Comparison rationale

Boundary and evidence layers

The comparison concerns accumulated graph content and its retained projections, configuration and recovery state. It does not classify the whole runtime's capabilities as memory. SQLite is the inspected persistence substrate; “knowledge graph” names the logical container. Disk snapshots and recovery files are additional retained objects. Ephemeral process caches and printed import proposals do not become independent memory stores merely because they have representations. See RT-OBJ-knowledge-graph, MEM-OBJ-snapshot-metadata and RT-RTE-git-import.

Implementation supports wired storage, structural admission, projection and recovery. Shipped prompts support claimed semantic curation and agent use. Neither layer establishes observed execution, dependable factual distillation or causal improvement. A wired database write affordance does not upgrade the journal/task semantic procedures to wired automation. See MEM-RTE-journal-curation, MEM-RTE-task-curation and RT-CLM-transaction-boundary.

Authority and admission are independent

Retained descriptions supply knowledge; retained plans and specifications supply directions. Acceptance criteria supply a prompt-level validation standard, whereas retained schema and predecessor state participate in executable validation and vetoes. Task status/field presence routes guide selection. Experience consumed to revise a durable plan has learning authority at that update consumer, independently of manual admission and without proving improved capacity. See MEM-BAP-journal-guidance, MEM-OBJ-task-specification, RT-RTE-task-guide, EPI-RTE-journal-review and EPI-RTE-undo-redo.

User-approved summary/design writes are manual even when models produce their wording. Conversely, an agent-written compact conversation capture or investigation finding without a content approval step is automatic at the claimed layer. Generic CLI/MCP/HTTP callers remain unresolved: physical I/O, author labels and a permissive API do not identify the admission decision. See MEM-RTE-journal-curation, MEM-RTE-task-curation and RT-RTE-graph-update.

Curation, trace learning and later consumption

Distilling retained activity into summaries is consolidation even when raw logs survive. Requested duplicate merges, retained-field revision and explicit cancellation/drop support separate operations, not a blanket synthesis classification. No recorded concrete generated proposition establishes the non-entailment needed to classify synthesis. The history squash implementation supports consolidation at a different layer from prompt-mediated journal summaries; it does not establish preservation of every provenance field. See MEM-OBJ-journal-summary, MEM-RTE-task-curation, MEM-RTE-history-maintenance and EPI-RTE-journal-review.

The positive trace-learning units are narrower than all journal/task writes. Automatic conversational captures become durable context for later review or planning. Investigation retains execution-informed diagnosis for subsequent task rereads. These qualify at the claimed layer without requiring new knowledge, observed improvement or weight updates. The manual day/period summary paths are not silently relabeled automatic. The Git importer remains a proposal producer, not an automatic durable learning route. See MEM-OBJ-journal-log, MEM-RTE-journal-curation, EPI-RTE-task-findings and RT-RTE-git-import.

Requested skill context is pull. Subscribed hook delivery and automatic config-directed projection supply non-model push paths. Their coarse event, query and availability filtering does not establish targeted identity matching or semantic ranking. An identifier on a requested record, output path or metadata row is insufficient. Opaque extra journal injection remains unresolved rather than making the system pull-only. See MEM-RTE-selected-read, RT-RTE-extension-load, MEM-RTE-document-projection and MEM-RTE-journal-curation.

Canonical records and coverage limits

Reconciliation supersedes the epistemic journal-summary object with MEM-OBJ-journal-summary. All summary classifications use that canonical object; the supersession changes no value or evidence layer. The broader task grouping EPI-OBJ-task remains valid for status and links alongside the finer findings/specification parts. History-maintenance and repair, projection and ingress, and journal-curation and journal-review remain distinct overlapping routes, not duplicate witnesses to be promoted in strength.

The records leave original live-content admission/provenance, extra calendar/health/activity payloads and parts of app/browser or remaining guide behavior unresolved. The workflow member extends guide inspection but does not supply a complete new retained-object or trace-input inventory. These gaps prevent complete-value assessments; they do not erase supported routes or justify a system-wide negative. See MEM-RTE-journal-curation, MEM-RTE-task-curation, MEM-RTE-selected-read and RT-RTE-task-guide.

Source-read log: none. This classification used only the supplied accepted members and reconciliation; no target-source ambiguity required a direct read.