Link memory profile

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

Comparison rationale

Boundary and coverage

The comparison covers accumulated content and its access, control and recovery parts, not everything Link stores or ships. The operative inventory comes from MEM-OBJ-memory-claim, MEM-OBJ-memory-controls, MEM-OBJ-retrieval-context, MEM-OBJ-session-capture, MEM-OBJ-wiki-pages, MEM-OBJ-wiki-cache, MEM-OBJ-semantic-index, MEM-OBJ-handoff, MEM-OBJ-usage-ledger and the cited recovery, synchronization, ownership and export routes. Their records retain the source passages; this profile adds neither records nor quotations.

Every axis is partial because the accepted inventory retains local gaps, notably detailed digest/auxiliary transformations and consumers, generic caller control and mixed initial provenance. These gaps do not suppress classified parts. Missing host traces limit operation and benefit, not the wired architectural findings. Static steering and external provider weights are not use-derived memory; the numerical vector cache is. Unpinned provider provenance remains a limit on its known derivation rather than a claim of trained Link parameters. See RT-OBJ-link-instructions, RT-CMP-model2vec, RT-CMP-fastembed and MEM-OBJ-semantic-index.

Forms, lineage and consumers

Markdown and JSON establish files, not graph/vector/KV databases merely because they encode links, dense vectors or keyed records. SQLite FTS and process maps have their own substrates. Graph access state is symbolic; dense distributed vectors are parametric; returned claim/article text is natural language. The same text can provide knowledge to an agent and ranking input to fixed retrieval code. Retrieval-only context has ranking force even though slim delivery omits it. See MEM-OBJ-wiki-cache, MEM-OBJ-retrieval-context and RT-RTE-context-recall.

Authored, imported, trace-extracted and other-compiled findings refer to separate paths, not exclusive labels for the whole corpus. Known extraction or rendering does not authenticate source strings or resolve initial content provenance. Structured body escaping is a documented encoding change, not an extra learned claim. See EPI-RTE-export-rendering.

Lifecycle controls enforce focused selection only. Generic wiki delivery remains an independent consumer of still-present archived, expired or other-project pages; its compaction omits memory-specific warnings. Thus invalidate is bounded withdrawal, not universal nondelivery. See MEM-RTE-wiki-read and EPI-RTE-memory-withdrawal. Dismissal membership validates candidates and enforces a veto; ownership hashes validate replacement and enforce its managed-output gate. Neither supplies claim truth. Journal state routes recovery rather than becoming factual advice.

Admission and maintenance

Manual findings describe explicit human decisions requested by the contract, not a verified universal approval gate. Generic callers, Boolean confirmation and source strings cannot establish actual human control. Software/model content admission within an initiated workflow remains automatic. Those findings do not upgrade unknown explicit replacement mechanisms. See RT-RTE-memory-admission, EPI-RTE-review-marking, RT-RTE-agent-wiki and RT-RTE-local-recovery.

The read-only planner is not implemented semantic consolidation. Known curation instead includes transcript reduction, actual capture deduplication, append and metadata revision, branch-local withdrawal, deletion/FIFO pruning and the reviewed-state ranking boost. Requested external-agent synthesis retains claimed basis. A proposal-normalization fidelity defect does not turn consolidation into synthesis. Query excerpt construction creates no retained successor and is not another durable consolidation write. See MEM-RTE-consolidation-plan, MEM-RTE-memory-lifecycle, MEM-RTE-capture-cleanup, EPI-RTE-proposal-normalization and EPI-RTE-context-compaction.

Read-back and trace-fed writes

Explicit requested recall is pull even when software selects/ranks its response. Startup, first-response autobrief and prompt reminders independently supply memory and are push. Within those push chains, coarse eligibility and budget rules are distinct from actual project matches and guard lexical overlap. Identifiers on requested files are not counted as identifier push. Automatic loads of retained access/control state are named separately from requested content delivery. See RT-RTE-session-delivery, MEM-RTE-guard-reminder and MEM-RTE-usage-feedback.

Automatic conversational extraction produces retained notes/candidates consumed by later review/startup routes; final claim acceptance is a separate mechanism. Usage events automatically produce durable history that changes later cooldown choices, without recording whether the model relied on returned content. Operation activity automatically produces durable markers/manifests that choose later recovery actions. These qualifying trace-fed writes have session-logs or event-streams as their original sources. Merely recording a tool-return identity does not establish consumption of a tool trace's request/result body. Mixed imports and optional project history retain unknown origins; no tool-traces or trajectories finding is invented from a label. No system-wide trace-learning negative follows from fixed provider parameters, read-only plans or nonadopted salience. See RT-RTE-session-capture, MEM-OBJ-session-capture, MEM-OBJ-usage-ledger, MEM-RTE-usage-feedback, RT-OBJ-operation-journal and RT-RTE-local-recovery.

These profile findings establish automatic durable behavioral-context/control paths, not improved future capacity, exercised self-improvement or causal benefit. Learning authority at the capture-to-memory update consumer is distinct from knowledge/instruction authority at the later host consumer.

Canonical records and source reads

The supplied reconciliation supersedes EPI-OBJ-wiki-page with MEM-OBJ-wiki-pages, and EPI-RTE-consolidation-plan with MEM-RTE-consolidation-plan. All classifications use these canonical destinations. The memory-page container and its claim/control/context parts remain valid distinct scopes; no invented split supersession is used. Shared checker and post-authoring completion routes remain distinct; neither converts static validation code into accumulated memory.

Source reads: none. All classifications derive from the supplied accepted records and their annotations, resolved through the supplied reconciliation.