Auto-claude-code-research-in-sleep memory profile

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

Comparison rationale

The inventory remains partial on every axis. The accepted records distinguish inspected helper behavior, model-followed procedures, opaque conversation state and included but uninspected maintenance. No supplied execution or causal evidence upgrades a finding to observed or causally supported. Local known units mean classification coverage, not reliable model compliance.

Canonical identities and scope

Reconciliation supersedes the epistemic query-pack object with MEM-OBJ-query-pack, and the epistemic wiki-claim object with MEM-OBJ-claims. All classifications use those canonical objects. The separately retained graph is MEM-OBJ-edges, not part of a claim page. Construction and consumption remain separate routes: EPI-RTE-query-pack supports wired compression, while EPI-RTE-query-consumption supports claimed later use. No supersession promotes either status.

The wiki graph uses file-backed triples. Its logical graph organization does not establish a distinct graph storage service. Provider weights are inference components, not retained memory updated by these routes. Opaque conversation state remains included, but a returned final message cannot classify its internal substrate or encoding. See MEM-OBJ-edges and RT-RTE-codex-call.

Transformations and authority

The maintenance storage and form branches keep separate evidence bases. Proposal/ledger production and skill replacement specify files through claimed procedures; only receipt stamping supplies wired storage on its own branch. Staged unified-diff syntax is symbolic at a claimed procedural basis, while the receipt implementation supplies wired symbolic hash/identity fields. The receipt witness upgrades neither storage nor form on the proposal branch. See RT-OBJ-skill and RT-RTE-meta-landing.

Reviewer-derived suspicions and judgments do not establish consolidation. The accepted records specify optional updates and append/replay, but do not establish reduction of existing claims. This local curation operation remains not-determinable, independently of the supported consolidation of the wiki pack and compact handoffs. Trace-fed derivation and later consumption do not require a known curation operation. See MEM-OBJ-reviewer-memory and MEM-RTE-reviewer-memory.

Fixed excerpting is consolidation. Filename order and character budgets are not semantic ranking, while retained outcome fields do route inclusion. The later model's instructed banlist affects candidate priority, not a guaranteed filesystem or action veto. Graph evidence, claim scope and full experiment reasoning can remain stored without reaching ideation. These limits follow MEM-RTE-pack-read, MEM-CLM-pack-contract and MEM-BAP-wiki-ideation.

Structural checks and scanners are not themselves accumulated memory merely because they control a write. By contrast, retained provenance receipt fields/hash are consumed by the eligibility checker and thus carry validation and enforcement locally. The receipt proves neither correctness nor reviewer execution. Full review trace writer/native-validator internals remain unknown. See RT-RTE-meta-landing and MEM-OBJ-review-traces.

Overwrite supports evolution; a retraction status supports requested withdrawal, not removal of stale graph edges. Primary-key skips alone supply no dedup finding. Novel prose supplies no synthesis finding without recorded non-entailment. Requested stage-summary synthesis also cannot be promoted into a verified curation mechanism merely from its name. See MEM-RTE-evaluation-write, EPI-RTE-wiki-retention, EPI-RTE-result-judgment and RT-RTE-research-pipeline.

Read-back and trace-fed updates

A requested wiki pack is pull even though a mutation automatically rebuilds the file. Next-round full-memory injection and automatic compact-context substitution are push operations with coarse selectors. Their selected file names alone do not establish identifier-based targeting. Thread continuation is explicitly requested pull; internal provider selection stays uninspected. See MEM-RTE-pack-read, MEM-RTE-reviewer-memory, MEM-RTE-compact-recovery and RT-RTE-codex-call.

Trace learning here means automatic trace-fed production of durable later context or guidance, not measured improved capacity. Compact experiment logs, reviewer-derived memory, evaluation-derived outcome guidance and the next-cycle maintenance ledger supply claimed positive chains. Verbatim logs alone do not. Human-selected live skill adoption is manual even though proposal/ledger production is computational. Experience supplies learning authority at an update consumer; derived context supplies knowledge at a later reviewer or ideator; revised live skill text supplies instruction. These are separate paths, not a blanket authority inference. See MEM-RTE-compact-recovery, MEM-RTE-reviewer-memory, EPI-RTE-result-integration, RT-RTE-meta-landing and RT-BAP-skill-instructions.

The remaining unknowns are consequential: findings production, opaque provider maintenance, global event consumers, trace serialization/native evidence checks, style extraction and platform alternatives prevent complete source, admission and consumer inventories. No local witness resolves them, and no trace-learning negative is inferred from ordinary recovery counters or manual skill landing. The accepted memory report retains these boundaries in MEM-OBJ-compact-files, MEM-OBJ-events and MEM-OBJ-review-traces, with the provider boundary in RT-RTE-codex-call.

Source-read log

No target source files were read for this classification. Only the supplied contracts, accepted members and reconciliation were used; no new records, evidence or quotations were added.