Autocontext memory profile

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

Comparison rationale

This is a classification of accepted records, not a fresh source audit. All positive implementation findings are architectural wiring unless their local basis says afforded. No record supplies an observed run or causal experiment for these memory paths. The profile makes no claim of successful learning, reliable advice, safe generated validators or exercised self-improvement.

Boundary and part distinctions

MEM-OBJ-context-bundle is a valid mixed container. Its eight parts stay separate where their consumers differ. Advice and four mutation text bodies provide instructions through model interpretation. In particular, the context-policy renderer does not implement a programmatic context budget, and a completion-check string does not execute a veto. MEM-OBJ-routing-config selects actual runtime settings. MEM-OBJ-harness-validator supplies conditional executable validity criteria and an enforced refusal path. MEM-OBJ-tool-spec has symbolic source syntax but a noncallable, model-interpreted reference consumer. The JSON wrapper never gives all these parts the same force.

MEM-OBJ-generation-artifacts distinguishes raw replay/metric records from derived lessons. MEM-OBJ-compaction-entry distinguishes a retained audit summary and lookup metadata from the compacted prompt body and from opaque host state. Its Pi inspection consumer supports requested reference delivery, not automatic continuation from a complete checkpoint. RT-CMP-training-checkpoints supports parametric form and file-backed checkpoint storage; a numerical form does not require guessing a separately hosted storage service. The source pin does not pin exact runtime weights.

Supersessions and admission

Reconciliation supersedes the epistemic lifecycle and import route declarations with MEM-RTE-lesson-lifecycle and MEM-RTE-package-import. All profile route citations use those canonical records. The epistemic payload objects remain distinct and support imported lineage, form and attributed-history limits. Neither valid container is superseded by its parts. The TypeScript session extraction record overlaps the broader Python/TypeScript preview record without replacing it; neither establishes durable promotion.

Software/model admission is automatic even when a human starts a run. Optional approval methods in RT-RTE-npm-generation, generic lifecycle callers and MCP/library package callers do not establish human agency. The separately described operator CLI package-selection branch affords manual admission. That weak, local warrant is not upgraded by the stronger automatic implementation witnesses. Unknown initial authorship does not erase known automatic bootstrap and subsequent candidate admission.

Curation and read-back

MEM-RTE-component-compaction implements reduction of existing material. Loss of a scope condition is a fidelity risk, not evidence that the operation creates a new claim. MEM-RTE-lesson-maintenance implements exact deduplication and replacement; curator-mediated semantic consolidation is afforded rather than asserted as an observed faithful transformation. The controlled synthesis value is not established merely by a request for rules or by failure to prove semantic preservation. EPI-RTE-role-proposals keeps that relation unresolved.

Generation-triggered selection in RT-RTE-context-assembly is push relative to the role consumer. The actual scenario/run/digest identity selects delivered parts; coarse ablation, availability, freshness and budget filters also act. The compaction keyword preference adds lexical selection without embedding inference. In contrast, MEM-RTE-requested-readback fulfills a consumer request. Its requested identifier and lexical search do not create extra push findings. MEM-CLM-file-consumption affords a later file-reading role, not host injection merely because skill symlinks exist.

Trace-fed updates and remaining coverage

The explicit durable witness is MEM-RTE-lesson-maintenance: attempt experience becomes retained warning/lesson content and later role prompts receive it. Existing lessons and score trajectory also feed durable curator replacement, giving retained experience learning authority at the updater and instructions downstream. The workflow's annotation on RT-RTE-scenario-generation places support roles before the current tournament. The warning produced at persistence still consumes current results; no claim that the coach saw that future tournament is necessary.

RT-RTE-training-experiments and RT-RTE-kernel-evolution also connect recorded experience, automatic retained source updates and later proposal/execution consumers. This supports local trace-learning findings without assuming all checkpoints are trained from session logs. Benchmark/backend execution feedback supplies tool traces, and linked candidate/action/outcome histories supply trajectories. No original event-stream or conversational-session source is inferred from an adapter name or from a report's existence.

Every axis remains partial because the accepted inventory includes unresolved alternatives. The records retain uninspected notebook/report writers, hint internals, evidence processing, npm maintenance/memory packs, production-trace/dataset curation, session/mission reload, configured hooks and opaque provider/host state. Some loaders and campaign consumers are known, but their complete origin, admission, storage or later-consumer chains are not. These gaps prevent complete-value statistics, system-wide negatives and a universal admission or revocation guarantee. They are inspection limits, not evidence of absent memory.

Source-read log: none. This job used only the supplied member/type/contract inputs and reconciliation; it added no source evidence or declarations.