Company Brain memory profile

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

Comparison rationale

This profile classifies accepted records, not new source observations. The scoped inventory is the union of the material memory mechanisms recorded in the supplied reports. Every axis retains unresolved service state and relevant available-but-uninspected alternatives instead of treating an inspected local route as a complete system inventory. All positive implementation findings concern wiring or affordances. There is no supplied observed run or causal experiment.

Objects and provider limits

A submitted document, the hosted derived text, local access metadata and a later rendered summary are different operative parts. MEM-OBJ-brain-document and MEM-OBJ-derived-memory support service-object storage at the application interface. They do not establish vector storage or a parametric representation of an embedding. SQL-backed retained objects support rdbms without assuming an unrecorded SQLite engine. The workspace-prompt and schedule records do not identify their physical persistence implementation, so those units remain unresolved even though persistence itself is wired.

Extraction policy has a known local provisioning route but an opaque later remote consumer. Its instruction authority and semantic curation requests retain claimed bases. MEM-CLM-service-curation supports requested duplicate merging, supersession, transient forgetting and pattern inference; none becomes a wired remote transformation from the strength of local submission or liveness filtering. Topic splitting changes access organization and same-text metadata, not necessarily one of the controlled claim-curation operations. MEM-RTE-node-splitting retains partial-update and tag-normalization limits.

Admission, authority and trace-fed reuse

Approval is specific to a mechanism. RT-RTE-instruction-admission establishes requester-approved skills and editor-supplied HTTP text; MEM-RTE-config-edit establishes a separate model-mediated admin writer without that per-text decision. Generated research drafts are automatically stored; reviewer replacement and explicit publication are separate human decisions under WF-RTE-draft-delivery. Generic debug caller identity does not resolve human control under MEM-RTE-debug-capture.

Knowledge delivered to an answering model, directions delivered as standing guidance, and retained experience used at a durable update consumer carry separate authority findings. Raw channel spools and carried style state can be learning inputs at distillation/observation, while the resulting content supplies knowledge or instruction downstream. Stored tags/counts/scores influence ranking; mappings/cursors/timing route operations; generation/status/liveness criteria validate and veto eligibility. No inference treats every descriptive tag as a factual warrant or every instruction as executable enforcement.

Automatic trace-fed document/style production, deterministic conversation compaction and checkpoint construction retain behavior-shaping content for later invocations. These satisfy trace_learning at the wired-capacity layer without weight changes, demonstrated improvement or a mandatory novel claim. Research output reuse is also retained history for later planning, but its complete source mixture and selection horizon remain partial. The schedule/configuration records do not by themselves settle qualifying trace provenance. A route not established as trace-fed is not a system-wide negative. No accepted bounded-absence record warrants a trace_learning no union.

Sources are classified by origin: Slack conversational messages are session logs, including when stored with timestamps or processed by an observer. Actual tool requests/results support tool-traces. The field called trajectory does not alone establish trajectories as linked state/action/outcome experience. Imported web information and static extraction instructions do not become event streams merely because software processes them.

Selection and chains

MEM-RTE-ambient-recall and MEM-RTE-style-recall independently supply context. Person/thread identity matching is targeted push selection; liveness, candidate caps, available baseline and recognized buckets are coarse selection. MEM-RTE-requested-recall, skill load, entity lookup and human readers fulfill requests, so their semantic/lexical query mechanisms do not automatically establish inferred push signals. Approval-triggered restoration fulfills a request for a specified continuation; independent checkpoint reload is a distinct push operation. The schedule callback independently starts execution, whereas a reviewer's draft-ID send requests delivery. Automatic history supply to planning remains separate from that requested send.

An inspected readable display does not resolve the opaque selected provider payload or embedding provenance. Restoring an investigation checkpoint establishes later state reuse, but MEM-RTE-investigation-recall leaves complete final rendering unresolved. Such local gaps survive positive findings.

Canonicalization and source-read log

The supplied reconciliation resolves three duplicate identities. Style-object classifications cite MEM-OBJ-agent-style, not the superseded epistemic style object. Partition classifications cite MEM-RTE-node-splitting, not the superseded epistemic partition route. Publication/editing classifications cite WF-RTE-draft-delivery, not the superseded epistemic sender. Those supersessions change citation identity, not the analysts' warrants. Distinct combined routes and their parts remain distinct.

Administrative reception acquisition/framing/interpretation under EPI-RTE-outcome-acquisition, EPI-RTE-outcome-frame and EPI-RTE-admin-consumption establishes an operative feedback consumer, but no automatic persisted criticism or planning-policy update. Its transient outcome payload is excluded from the retained-object inventory; retained sent references and body history remain included. This does not turn missing episodes into absent criticism or erase implemented feedback consumption.

Additional frozen-source reads: none. All classification warrants come from the supplied accepted records and their annotations; no source fact was added or report verification reopened.