Pal memory profile

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

Comparison rationale

This profile classifies accepted records only. It adds no source quotations, records or execution evidence. No source files were read for this job, so there is no additional source-read log. Configuration and implementation support wiring; prompts support requested transformations. Neither establishes actual admission, delivered payloads, changed decisions or benefit.

Inventory and evidence layers

The memory boundary includes content and its access structures. Raw files, articles, edited guidance and saved products are not interchangeable with manifest flags, index descriptions or routing vectors. PostgreSQL storage is known even where framework payload form is opaque. Git is optional retained-file replication, not a copy of learning/session databases. These distinctions follow MEM-OBJ-raw, MEM-OBJ-wiki-articles, RT-OBJ-manifest, MEM-OBJ-pal-knowledge, MEM-OBJ-user-memory and RT-RTE-git-sync. The workflow's push-failure qualification changes confidence in remote delivery, not the existence of the repository substrate or local file retention.

The storage inventory can therefore be known while most other axes remain partial. In particular, the registered knowledge-service handles are included, but registration supplies neither a concrete additional consumer nor a fully specified admission contract. Framework learning/session payloads and possible checkpoints stay unresolved rather than being replaced by their intended readable descriptions. See MEM-CLM-service-boundary, MEM-RTE-learning-reuse and MEM-RTE-conversation-reuse.

Canonicalization

Reconciliation supersedes epistemic raw, lint-report, routing, learning and SQL object duplicates with their memory declarations, and the epistemic routing-write duplicate with MEM-RTE-knowledge-update. The profile uses those canonical destinations. This affects citation identity, not evidence strength or a value union. Wiki summaries and concepts remain valid separate parts of MEM-OBJ-wiki-articles; their unsuperseded EPI-OBJ-summaries and EPI-OBJ-concepts support local transformation distinctions. EPI-RTE-learning-adapt remains broader than operational-learning reuse because it also includes personal memory; it is not treated as a replacement for either memory route.

Non-obvious mappings

Dense retrieval representations establish parametric form and vector storage, not model-weight learning. Routing text and vectors have ranking force at search, while selected descriptions have routing force at the model's next retrieval choice. Manifest flags likewise guide compile selection but are not a truth-validation criterion. Retained correction labels and recency have requested ranking force; preference/voice and strategy text have intended instruction force. Experience at the update consumer has learning authority independently of that downstream instruction force. See RT-CMP-openai-embedder, MEM-OBJ-pal-knowledge, RT-OBJ-manifest, MEM-BAP-retained-guidance and EPI-RTE-learning-adapt.

Compiler summaries support claimed consolidation, and concept/learning revision instructions support claimed evolution. The semantic relation of an actual generated article remains indeterminate. Neither a request to synthesize nor missing evidence of entailment establishes creation of a new claim. Search-before-save is not an implemented duplicate merge; vector recreation is not by itself controlled semantic invalidation with history preserved. Lint advice is not enforced withdrawal. These limits follow EPI-RTE-compiler-transform, MEM-RTE-learning-reuse, RT-RTE-context-refresh and MEM-OBJ-lint-report.

Read-back distinguishes requests from automatic supply. A model-requested file fetch remains pull even when index descriptions helped choose its path. Automatic history and learning additions are push. Recent-history run limits support a coarse window component, but neither file paths nor past-session IDs establish an inspected identifier-matched push. The automatic learning selector is unresolved despite the configured hybrid store. See RT-RTE-navigator-work, MEM-RTE-knowledge-update, MEM-RTE-learning-reuse and MEM-RTE-conversation-reuse.

Trace-learning positives belong to automatic interaction-to-durable-guidance chains, not raw transcript retention, source compilation in general or provider weights. The session-origin path is supported; exact tool-trace inclusion and other source mixtures remain unknown. No local or global no is invented without a bounded absence record. Unresolved service writes and checkpoints also prevent trace-source inapplicability. See MEM-RTE-learning-reuse, MEM-RTE-conversation-reuse, MEM-OBJ-session-history and MEM-CLM-service-boundary.

Automatic model/tool admission is distinct from human authorship. Explicit user editing affords manual admission; generic API submission and remote Git authorship do not identify a human decision. Running a loader or evaluator does not make each metadata insertion or file removal manual. These distinctions preserve known mechanisms without assigning unknown initial or alternative content control. See MEM-OBJ-user-files, RT-RTE-raw-ingest, RT-RTE-git-sync, RT-RTE-context-refresh and RT-RTE-evaluation-driver.