Synapptic memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison follows retained parts and actual consumers, not file formats or the enclosing model runtime. Profiles and narratives carry natural-language content; weights, IDs, exclusions and cache keys carry symbolic control. Relay tables and tracker state have operator/bookkeeping consumers, not an established instruction-learning path. The opaque resumed condition session remains included even where its form, origin and transformations cannot be classified. See MEM-OBJ-claude-session, MEM-OBJ-request-log and MEM-OBJ-session-bookkeeping.
Reconciliation supersedes the epistemic profile-text, profile-fields, archetype and benchmark-case objects with their memory counterparts. The profile therefore cites MEM-OBJ-profile-observations, MEM-OBJ-profile-metadata, MEM-OBJ-archetype and MEM-OBJ-benchmark-cases, not the superseded objects. Returned-observation and evaluator-result containers remain valid but are not counted as additional stored parts merely because another member describes them.
Learning authority at merge/synthesis/feedback consumers differs from downstream knowledge or instruction authority. Automatic admission is not human approval merely because an operator starts ingest or installs a hook. Benchmark verdict writes precede the separate human exclusion decision. A human-requested global snapshot replacement is classified at its admission mechanism rather than inferred from a read of history. See RT-RTE-ingest-profile, RT-RTE-benchmark-feedback and MEM-RTE-state-maintenance.
Trace-fed extraction qualifies even without demonstrated improved performance. Extracting standing observations consolidates retained conversation; similarity merging, longer-text replacement, decay and cross-project promotion have separate implemented effects. Neither a command named synthesis nor a requested failure explanation establishes a novel claim. Narrative semantic novelty therefore stays unresolved rather than adding the curation value synthesize. The omitted separate evidence field also prevents interpreting the compiled profile as a guaranteed warrant-preserving representation. See EPI-RTE-profile-curation, MEM-OBJ-profile-observations and RT-RTE-publish-archetypes.
Pull and push can occur within one chain: a driver requests files, then automatically supplies selected retained content to a model. Browser FTS selection fulfills an operator request; it is not lexical push into an agent. Target-model identity matching actually selects the verdict field used by synthesis, whereas integration merely computes filtering counts before writing unchanged saved prose. This mapping does not claim model-specific publication filtering. Host loading is only afforded, so its push/instruction findings retain that basis rather than inheriting local writer wiring. See MEM-RTE-browser-readback, EPI-RTE-narrative-inputs, RT-CLM-target-filtering and RT-BAP-archetype-host.
The benchmark also has a distinct lexical push selector: at per-guard setup, chosen guard text selects a matching block of the retained full archetype for removal from WITHOUT, or the guard is appended to WITH when absent. This controls which narrative parts the respondent receives without a respondent retrieval request. It supports inferred-lexical at wired basis, separately from coarse treatment choice and identifier-based cached-case reuse. Heuristic matching limits comparison fidelity; it does not erase the implemented selector. See RT-RTE-benchmark-feedback and EPI-RTE-benchmark-judgment.
The session conversation route supports session-logs as an original trace source. The SessionEnd event is a trigger, not evidence that event content feeds the qualifying write. Benchmark feedback qualifies as automatic activity-fed durable adaptation, but its synthetic inference responses and judgments do not unambiguously establish one controlled source class in the accepted records. The source axis retains that gap instead of treating every API result as tool-traces. No bounded absence record licenses a system-wide trace-learning no or an inapplicable trace-source axis. See RT-RTE-session-hook, RT-RTE-benchmark-feedback and MEM-OBJ-benchmark-results.
All positive findings are implementation wiring or explicitly afforded external consumption. No observed or causal basis is supplied. Partial assessments preserve missing facts about opaque context and incompletely inspected live UI paths; they do not turn those paths into absence. No source files were read for this classification: the supplied accepted records and reconciliation were sufficient, and no new quotes or records were added.