Tardigrade memory profile

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

Comparison rationale

This profile classifies accepted records, not new source inspection. All positive implementation findings are wired, not observed activation or benefit. The optimizer findings remain claimed because the shipped procedure, not an executed optimizer, is the evidence. Every axis is partial: positive witnesses coexist with the named opaque/custom or uninspected alternatives. The scope is accumulated memory and later consumers, not every storage or permission feature of the runtime.

Canonical identities and parts

Reconciliation supersedes the epistemic workspace object/route with MEM-OBJ-workspace-text and MEM-RTE-text-workspace, and the epistemic current-summary route with MEM-RTE-atom-compaction. All classifications use those canonical destinations. No value changes merely because an alias is superseded. The summary text part EPI-OBJ-summary-text remains distinct from MEM-OBJ-compaction-state, which also owns selector/control fields. Likewise, continuation and effect-checkpoint parts do not replace their journal container. Current code-mode recovery is not the legacy spill implementation; see RT-RTE-code-execution and MEM-RTE-spill-readback.

Consequential mappings

Automatic summaries qualify for trace learning because the trace-to-retained- summary-to-later-inference chain is recorded, even within a run. Consolidation classifies the implemented compression purpose; unverified faithfulness neither proves synthesis nor defeats that classification. Prior experience supplies learning authority at the summarizer, while summary text supplies knowledge at later inference. The summary's user-role delivery does not alone make it standing instruction. See MEM-RTE-atom-compaction and MEM-RTE-legacy-compaction.

Retained rejection errors form a separate automatic, within-turn guidance route: EPI-RTE-response-check. This route qualifies through retained rejection state and its later corrective system consumer, not through an inferred standing actor rewrite. Its validator-feedback source is left partially classified rather than called an environmental event merely because it is logged. The default Bun example has no output contract, so this conditional route is not asserted active in every invocation.

Read-back units distinguish requested workspace/event retrieval from automatic context, identity-matched continuation/file delivery and recovery selection. Substring grep on a requested spill query is pull, not inferred-lexical push. Coarse size/availability windows coexist with identity matching of actual retained parts. Compaction's model-generated text does not establish a judgment-based retrieval selector. See MEM-RTE-content-replay, MEM-RTE-text-workspace, MEM-RTE-event-inspection and RT-CLM-request-before-execution.

The prescribed optimization loop adds trace-fed authored revisions, candidate score ranking, criteria-based validation and promotion only at its claimed layer. Its installed executable artifact route is independently wired; that witness does not upgrade procedural selection or prove a teacher generated or installed anything. User authorization of scope is not per-candidate manual admission. Generic workspace, artifact and state-adoption callers retain unresolved content control. See RT-RTE-optimization-guidance, RT-RTE-artifact-admission and MEM-RTE-state-adoption.

Coverage limits and read log

The opaque continuation's local serialization/storage is classified without pretending its display reasoning describes every consumed part. File refs are symbolic; their bytes and custom backend substrates remain unresolved. Raw spill retention and checkpoint encoding do not alone establish trace learning. Repair projection withdraws current reliance while keeping raw history, without proving rejected assertions false. These limits follow MEM-OBJ-provider-continuation, MEM-OBJ-content-parts, MEM-OBJ-effect-checkpoint and MEM-RTE-legacy-compaction.

No source files were read for this classification; only the supplied members and governing contracts were read. No new evidence, quotations, records or target execution were added. Missing execution evidence prevents observed/causal bases, not the recorded wired classifications. No bounded absence record warrants a system-wide trace-learning no, and unresolved alternatives cannot be converted to inapplicability. The advertised model log reload, custom backends, attachment provenance, remaining spill producers and maintenance automation prevent stronger coverage assessments; known routes do not erase those gaps.