Tracecraft memory profile

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

Comparison rationale

The profile classifies accepted records, not additional source evidence. Bucket-backed content is service-object, while the local cursor and durable trial output are files. Host SQLite is an acquisition source. Shipped Markdown instructions and local credentials are outside accumulated memory, so their instruction authority and access controls do not become memory findings. The object inventory is supplied by MEM-OBJ-session-parts, MEM-OBJ-mirror-state, MEM-OBJ-agent-registration and MEM-RTE-evaluator-readback, together with the other cited object records.

Software storage I/O is not admission agency. Caller-submitted values, completion notes and uploaded files retain unresolved human control. In contrast, mirror deltas, cursor advancement, compiled compact output and evaluator row emission are automatic after initiation. See RT-RTE-memory-write, RT-RTE-session-mirror, MEM-OBJ-inbox-cursor and MEM-RTE-evaluator-readback.

Claim ownership supplies both an admissibility criterion and a veto at completion. Status supplies enforced dependency readiness and routes the waiter; neither checks work correctness. Message/value reference use is only afforded at the external agent. An uploaded file cannot be classified by its filename, and a stored handoff directive cannot be assigned operative instruction authority without its later consumer. See RT-RTE-step-control, RT-BAP-step-polling, MEM-OBJ-artifact-file and MEM-OBJ-handoff.

Retained session metadata has two distinct effects. List/show/status views display inspection information, supporting knowledge authority. Session list also sorts by recency or size and applies a default limit of 20, so retained metadata supplies ranking inputs even though fixed code implements the ordering. The ranking finding is wired, not evidence of actual agent uptake or a learned scoring policy. Registration display remains separate from this sorting effect. See MEM-OBJ-session-meta, RT-RTE-session-mirror and MEM-OBJ-agent-registration.

Replacement revises retained entries. Deleting direct mail forgets those records. Cursor filtering suppresses delivery without deleting content; it is not retained-content deduplication. Transcript compact combines bytes and reduces object count, not the claims represented. Neither its name nor regex redaction establishes semantic consolidation or synthesis. See RT-RTE-mail-exchange, RT-RTE-session-compact and EPI-RTE-regex-scrub.

All established content deliveries are requested. A new-mail cursor, an exact file name, or a tail budget does not turn a requested result into push. Missing handoff-reader wiring and uninspected durable evaluator report reads prevent a complete pull-only classification and system-wide read-back-signal inapplicability. See MEM-OBJ-handoff, RT-RTE-mail-exchange and MEM-RTE-evaluator-readback.

Raw/transformed session retention is not by itself trace learning. The records identify no derived continuation summary with a later guidance consumer; opaque already-derived host checkpoints do not supply their original derivation chain. Evaluator emission also lacks an inspected durable-output-to-policy consumer. No analyst declared an evidenced-absence record. Accordingly the trace axes retain uncertainty rather than manufacture a bounded no or declare trace source inapplicable. This does not suppress any accepted positive learning finding: none establishes the qualifying chain. See MEM-OBJ-session-parts, RT-RTE-session-mirror and MEM-RTE-evaluator-readback.

Reconciliation supersedes the epistemic and workflow handoff objects with MEM-OBJ-handoff, the epistemic transcript object with MEM-OBJ-session-parts, and the epistemic history-loss claim with MEM-CLM-history-immutability. The profile uses those canonical anchors; these identity amendments neither increase basis nor resolve consumer gaps. The two evaluator route records remain distinct.

Source-read log: none. Classification used the complete supplied members and contracts only; no frozen source ambiguity required a source read. Implementation findings are wired, external usage is afforded, and none is promoted to observed operation or causal benefit.