LLM Repo Wiki memory profile

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

Comparison rationale

This profile classifies accepted records, not a deployed database. All wired findings concern inspected architecture. No supplied run or experiment establishes activation, reliable summary preservation, factual accuracy or benefit. The inventory and its limits follow MEM-OBJ-chunks, MEM-OBJ-page-revisions, MEM-OBJ-project-context, MEM-OBJ-chat-messages, MEM-OBJ-chat-summary and MEM-OBJ-eval-results. Diagnostic results are included because citation evaluation really reloads prior pairs and judgments, not because every report becomes production memory.

The only reconciliation supersession replaces the epistemic lint route with MEM-RTE-lint-wiki. This profile uses that canonical destination. Object containers and projections remain distinct; no other supersession changes a value.

Database versus vector versus files. Dense vectors share Postgres rows with text, but their numerical form and cosine-ranking consumer warrant separate vector and parametric findings. Repository acquisition does not make the application's retained database a Git repository substrate. Exported Markdown and evaluator result files add files, not an inferred remote service store. See MEM-OBJ-chunks, MEM-OBJ-wiki-chunks, MEM-RTE-search-knowledge and RT-RTE-external-consumption.

Admission is not authorship. Automatic source writes, proposal generation, chat creation, model/tool/server message admission and summary admission coexist with afforded human supply of retained user turns, review/editing and selected tool approval. The human-facing chat path lets the user decide what text to supply; server persistence does not convert that admission decision to automatic control. A benchmark's human author label does not establish manual admission. Initial installed provenance and generic project API caller control remain unresolved, rather than erasing known subsequent mechanisms. See RT-RTE-chat-turn, RT-RTE-revision-admission and MEM-OBJ-project-context.

Read-back has several operations. The memory member's observation that routes are job/turn/request-triggered rules out an independent unsolicited daemon, not automatic push within context assembly. A requested search response is pull. A turn's automatic summary/project/history injection and code-selected evidence delivered to a separate writer are push under this profile's contract. Identity matching selects actual chat/wiki/page parts; budgets and availability are separate coarse filters. Author evidence selection uses embedding and lexical signals. No inferred-judgment signal is assigned merely because a model writes query strings or summary prose. See RT-RTE-context-fold, RT-RTE-author-pages, MEM-RTE-search-knowledge and MEM-BAP-project-brief.

Consolidation does not prove fidelity or novelty. Folding is implemented reduction of retained conversation into a durable summary. Its requested compression supports consolidate even though preservation is unchecked. Existing-page replacement supports evolve, rejection with surviving history supports invalidate, and explicit forgetting supports decay. Chunking/index replacement and deduplicating retrieval hits do not alone establish consolidate or dedup over retained memory. No concrete generated proposition establishes synthesize; the indeterminate semantic routes remain included, as RT-RTE-author-pages and EPI-OBJ-page-prose require.

Trace learning is a write-route classification, not measured improvement. Stored conversational turns and clipped tool requests/results feed automatic folding; its persisted output enters later chat instructions. This warrants a local wired "yes", session-logs and tool-traces. It does not warrant event-streams merely from log serialization, trajectories merely from sequential messages, or provider weight learning. The input's learning authority at the update consumer differs from summary knowledge/instruction authority at later chat. Diagnostic retention/resume has no established automatic production adoption and no canonical evidenced-absence record supporting "no"; those alternatives remain unresolved. The derived union therefore contains "yes" without a complete learning-inventory claim. See RT-RTE-context-fold, MEM-BAP-summary-context, MEM-OBJ-wiki-log and EPI-RTE-citation-evaluation.

Source-read log: none. Classification required no additional source reads; accepted member records supplied the warrants and retained passages.