Decapod memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory separates retained content from its access and admission metadata. It includes coordination and acceptance memory where later invocations consume it, but not the runtime's entire filesystem or external model weights. Static constitution delivery does not make shipped instructions accumulated memory. Live governance files and Git history are separate storage paths; the transient accepted-executor object is not an in-memory memory classification, although retained accepted-base Git authority is included (RT-OBJ-accepted-executor, MEM-OBJ-trajectory-artifact).
All positive findings are architectural, not observed operation or benefit. The inventory remains partial because the accepted reports explicitly retain untraced decide, federation graph, worker lesson, health/reflex, scheduled/evaluator, RPC-trace, spec initialization, full governance and remote-driver routes. The repeated unresolved unit keeps these included parts visible; its references identify adjacent accepted records, not nonexistent declarations for omitted mechanisms. No source fact outside those records establishes a value.
Reconciliation canonicalizes the duplicated knowledge, internalization manifest/payload, feedback and LCM-summary objects to memory declarations. It also canonicalizes internalization lifecycle and feedback proposal routes to memory declarations. These supersessions change citation destinations, not evidence strength or opaque-payload coverage. The epistemic promotion event, knowledge selection/withdrawal and evaluation parts remain distinct; LCM and knowledge routes that merely overlap remain separate. SQLite information from reconciled object declarations is preserved in the canonical object classification; the other recorded relational tables are classified as rdbms without claiming unacquired backend internals.
Requested returns remain pull even when fixed code chooses a prefix, ranks rows or performs budget checks. A requested archive ID, active prompt context, original hash or bundle path does not establish identifier-based push. Consequently read_back_signal has inapplicable local pull units but an uninspected axis: included alternative consumers prevent a completely assessed pull-only boundary (MEM-RTE-archive-context, MEM-RTE-aptitude-reminders, MEM-RTE-internalization-lease).
The LCM prefix qualifies as consolidation because code combines retained originals into a reduced representation of their existing content. The memory report's distinction from semantic summarization is retained: truncation may omit qualifications and unsafe slicing may fail. Neither defect implies synthesis. The qualifying interaction/tool branches also support trace_learning yes: automatic durable summary production and later requested context delivery suffice, without improvement evidence. Handoff has a separate trace-fed knowledge route, sourced from discrete coordination activity; its caller summary's additional provenance remains unresolved (MEM-RTE-lcm-summary, EPI-RTE-lcm-summary, MEM-RTE-handoff-memory). Learning authority at those update consumers and knowledge authority at later context consumers are separate findings.
The promotion event records an alleged approver and is later a procedural-admission criterion. It does not itself promote an existing knowledge row or authenticate human approval. Thus it supports validation/enforcement, not an automatic promote operation or manual agency. Library expiry/conflict/recency transformations retain afforded findings beside wired ordinary reads and promotion checks; no scheduled decay is asserted (EPI-RTE-promotion-record, MEM-RTE-promotion-lifecycle, EPI-RTE-knowledge-withdraw).
Opaque custom adapter bytes retain known file storage but unknown representation, derivation and inference effect. The noop alternative has empty content, while manifest and lease fields are symbolic with actual integrity/admission consumers. Neither a behavior-changing label nor a metadata attach supplies a parametric finding, a training route or an inference-context push (MEM-OBJ-adapter-payload, MEM-RTE-internalization-lease).
No new source reads were made. Classification used only supplied accepted reports, their annotations and reconciliation. No records, quotations or new evidence were added; no report verification was reopened.