Hermes Agent memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This is a revision-2 classification of the accepted records, not a new source account. Every axis is partial because the included unresolved alternatives differ by dimension. Local known units establish their scoped mechanisms, not complete system inventories. All positive mechanisms are wired unless a finding explicitly says afforded; none is observed or causally supported. The frozen analysis remains partial-loop inspection.
The reconciliation establishes no supersessions. The session envelope, its readable retained content, local summary and opaque checkpoint remain distinct parts. Notes/profile share a mechanism but not a file or budget. Skill content does not replace the package/control inventory. Review scheduling, candidate generation and guarded downstream retention are connected stages, not duplicate routes. These distinctions preserve the warrants of MEM-OBJ-session-content, MEM-OBJ-native-checkpoint, MEM-OBJ-skill-package and MEM-RTE-review-write.
The initial authorship of a store is independent of its later admission. Automatic model/software writes and explicit human pending approvals therefore coexist with unresolved initial files and learning-node UI control. The latter bypasses the optional tool gate; it cannot be silently assigned manual agency or excluded to make coverage complete. See EPI-RTE-memory-gate and MEM-RTE-pending-review.
Learning authority belongs to traces at update consumers; downstream notes, procedures and summaries separately provide knowledge or instruction. Local continuation summaries count as trace learning even when they merely compress existing claims. Saved goal feedback also changes later manager/repair behavior; its source is response history and execution feedback, not an event stream just because a gateway wraps a repair as an event. None of these classifications establishes semantic novelty, improved capacity or exercised self-improvement. See MEM-RTE-local-compaction, MEM-RTE-review-write and EPI-RTE-goal-disposition.
Skill archive/restore maps to decay, invalidation and promotion because it changes availability while retaining the package. Exact bullet reduction and tool-result deduplication supply dedup; extraction and summary reduction supply consolidation. Model prompts requesting class-level rules or optional curator consolidation do not by themselves establish synthesis. See MEM-RTE-skill-curation, EPI-RTE-recall-reshape and MEM-RTE-local-compaction.
A requested body or archive read remains pull, including its code-owned result delivery. Whole prompt blocks, summary selection and unsolicited provider prefetch are push. RetainDB cache consumption follows prior queued queries; ByteRover queries at prefetch time. Neither adapter exposes a sufficiently specific remote selector to assert inferred lexical, embedding or judgment selection. Identifier-bearing catalog entries and requested file paths are not identity-matched push evidence. See MEM-RTE-skill-cycle, MEM-RTE-session-recall and EPI-RTE-recall-consume.
Successful spill creation establishes file retention; the later parent read is afforded, not established as wired replay. The child summary's earlier schema verdict does not classify omitted or altered content. Raw retention alone is not trace learning. See WF-RTE-summary-return and WF-RTE-output-contract.
Provider ingress does not establish a trace-learning source. Queued turn pairs and explicit curate requests can carry traces, but the accepted records leave the derived transformation, admission and durable behavior-shaping result unresolved. The provider trace-source unit therefore has no positive findings and remains not-determinable, consistent with its trace-learning unit. This does not withdraw the supported sources of review, local handoffs or goal feedback. See MEM-RTE-provider-cycle and MEM-OBJ-provider-state.
No frozen target source files were read for this profile. All warrants come from the supplied accepted record passages. Opaque provider/checkpoint contents, frontend lifecycle equivalence, UI authorization/refresh, micro-summary commits, optional maintenance and alternate context/async consumers remain limits rather than inferred absences.