Phantom memory profile

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

Comparison rationale

The comparison classifies accepted records, not a newly inspected source. All positive findings describe wiring or affordances; no supplied observed run or causal experiment establishes activation or benefit. Axis coverage is partial because the included inventory retains SDK-owned conversation/compaction state, unknown initial producers and caller-dependent branches alongside known host mechanisms. A local known unit is complete only for its named path and parts.

Derivation and canonical identity

The reconciliation supersedes the epistemic episode, fact and procedure objects with MEM-OBJ-episodes, MEM-OBJ-semantic-facts and MEM-OBJ-procedures, and the epistemic extraction route with MEM-RTE-vector-extraction. Those canonical declarations supply the classifications. The separate epistemic retention, similarity and validity stages remain distinct; none is superseded by extraction. In particular, EPI-RTE-validity-change establishes withdrawal from default recall, not truth acceptance or universal erasure.

Compaction is a change action under ok, not a separate accepted status. RT-RTE-reflection-admission therefore supports changed-ok refresh for admitted compactions. It does not imply delivery of every admitted file: MEM-OBJ-evolved-files and MEM-OBJ-reflection-context distinguish fixed-loader/formatter delivery from reflection-only material requiring promotion. Reconciliation's residual conflict concerns memory's stale attribution of earlier runtime claims, not these current mechanisms. This profile neither reopens verification nor treats those requested corrections as outstanding.

Independent forms, forces and evidence layers

Dense access vectors are parametric representations in a vector substrate; sparse token features are symbolic. Neither is a Phantom weight-update route. MEM-OBJ-retrieval-indexes leaves historical embedding provenance unresolved. File prose, control metadata and binary attachments are likewise separate units: a display description cannot classify an opaque payload.

Retained text and access features can rank retrieval even with fixed scoring code. Fact validity can enforce exclusion; session/queue/job metadata can route later work. At reflection's update consumer, queued experience has learning force, while selected output files provide instruction or knowledge at the foreground model. These are separate paths, not an implication from the trace-learning label. The relevant routes are MEM-RTE-vector-context, RT-RTE-reflection-admission and WF-RTE-job-creation.

Known automatic admission never implies manual approval merely because the operator initiated work. Conversely, the authenticated caller on MEM-RTE-operator-memory and RT-RTE-operator-config does not establish human control of each body. Reflection/foreground consolidation, deduplication and promotion remain afforded where the records establish model guidance and tools rather than a particular performed transformation. A principle label, inferred preference or faithfulness defect does not establish synthesis.

Read-back and qualifying trace routes

Automatic vector selection and host continuity are push; requested memory tools, transcript search and file reads are pull. A requested lexical search adds no push lexical signal. Fixed-file inclusion is coarse, not identifier targeting merely because a filename exists. Session/tool identity matching in RT-RTE-chat-context does select retained parts; its caps and availability filters remain coarse. The evolution gate's content-sensitive judgment is an upstream selection into retained evidence for later reflection, independently of cadence's coarse batching and its fail-open admission alternative.

Automatic extraction and durable event-derived timeline summaries qualify as trace learning when their retained outputs reach later context, even without a novel claim or measured benefit. Raw chat persistence alone is not that update. MEM-OBJ-chat-timelines and RT-RTE-chat-context close the derived-summary and later-consumer path; compact-boundary counters do not close the opaque SDK summary path. The local procedure no is bounded by MEM-ABS-procedure-learning, not extended to initial/external procedures or the whole system. Original conversational turns remain session logs even when staged as JSONL. Actual translated runtime and tool events feeding timelines support event-streams and tool-traces. No record establishes linked state/action/outcome trajectories as an additional qualifying source.

Source-read log

No frozen-source files were read for this classification. Only the supplied contracts, boundary and accepted members were read; all warrants stay in their canonical records. No new evidence or analyst records are declared here.