WeKnora memory profile

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

Comparison rationale

This profile classifies accepted implementation records, not observed operation. All positive implementation findings remain wired; the explicit synthesis request retains claimed. No execution or benefit is inferred from retention, prompt delivery, successful validation or an available consumer.

Canonical records and scope

Reconciliation supersedes the epistemic chunk, memory-item and wiki-page objects with MEM-OBJ-chunk, MEM-OBJ-memory-item and MEM-OBJ-wiki-page. Their canonical IDs supply this profile's object citations. This changes no classification warrant: epistemic route records still describe distinct extraction, checking and maintenance functions. Context rewrite is not personal-store review; a checkpoint object is not its transformation route. Neither containment nor shared conversational origin merges these objects.

Every axis is partial over the stated inventory. Accepted records establish positive routes but retain uninspected alternatives and backend/payload gaps. The limits on MEM-OBJ-chunk, MEM-RTE-wiki-update, RT-RTE-sandbox-execution, RT-RTE-mcp-call and RT-RTE-skill-install prevent extending one inspected path to graph, multimodal, workspace, approval or host alternatives. A provider-backed image has a known exposed service-object boundary without a known complete payload. A diagnostic report does not classify that payload. Repository persistence without a pinned backend is retained as uncertainty, not silently projected to SQLite or a particular relational engine.

Consumer-relative mappings

Personal notes supply background knowledge, not permission or automatically binding instructions. Retained words/vectors also rank selection, and familiarity counts change executable rerank order. Those forces coexist with learning authority at the experience-fed updater. These are separate consumers, not a blanket consequence of trace learning (MEM-RTE-personal-recall, MEM-RTE-retrieval-conditioning). Checkpoint boundary fields route replay while checkpoint text supplies context. Skill documentation instead supplies standing instructions; a snapshot pointer routes later workspace resolution. Executable capability is not forced into a new authority category absent from the controlled vocabulary (RT-RTE-context-rewrite, RT-BAP-skill-image).

Tombstones supply both a candidate criterion and an enforced veto. Existing wiki row identities likewise supply validation and bounded enforcement at the rewrite gate. Neither is truth certification. Status exclusion and archival withdraw reliance locally; deletion does not imply retained history or global erasure (MEM-OBJ-memory-tombstone, EPI-RTE-wiki-guard, EPI-RTE-wiki-maintenance).

Human-capable APIs do not establish manual write agency. Explicit/manual origin labels and authenticated callers leave the actual content decision unresolved. This does not weaken known software admission of unflagged extraction, consolidation or snapshots. Conversely, a user initiating those workflows does not make each admitted output manually approved (MEM-RTE-personal-write, MEM-RTE-memory-review, RT-RTE-skill-install).

Transformations, selection and traces

Consolidation is supported for standing-note extraction, selected duplicate merges and generated continuation summaries. Raw checkpoint fallback has its own path: retention and later consumption qualify as trace-fed continuation, but copying an archive does not establish consolidation. The extraction prompt's deliberately non-entailed inference example supports a claimed synthesis class, not a wired guarantee that outputs contain new claims. Wiki rewriting supports evolution; indeterminate semantics do not prove synthesis (MEM-RTE-personal-write, MEM-RTE-memory-review, EPI-RTE-memory-extraction, RT-RTE-context-rewrite, MEM-RTE-wiki-update).

Requested search/read returns are pull. Turn-time recall, history supply and maintenance selection are push at their respective consumers. A document-ID match selects familiarity counts; a snapshot pointer selects the delivered image. These actual identity matches establish identifier signals. A slug on a requested wiki read does not. Coarse availability/budget selection remains separate from query lexical/vector ranking and maintenance-model judgment (MEM-RTE-memory-search, MEM-RTE-personal-recall, MEM-RTE-retrieval-conditioning, MEM-RTE-memory-review, RT-BAP-skill-image).

User-turn extraction, recurrence, familiarity feedback and durable continuation checkpoints have qualifying automatic trace-to-later-consumer chains. Their wired yes does not claim better future action. Q&A indexing proves later message-search output but not the further behavior-shaping consumer needed for a positive here. Installation repair preserves effects and selected gate findings, not replayed prior conversation; a diagnostic transcript alone cannot establish trace learning. These unresolved routes remain unresolved rather than receiving local no or inapplicable source labels (MEM-RTE-chat-index, RT-RTE-skill-install).

The source classification follows original content. Personal extraction uses user session lines. Familiarity uses answer-associated references, not user approval. Checkpoint history also includes AgentSteps and tool-call/result material. Neither a timestamp nor being handled by a completion callback adds event-stream provenance; step sequences alone do not establish trajectories (MEM-RTE-personal-write, MEM-OBJ-doc-affinity, RT-RTE-context-rewrite, RT-RTE-agent-turn).

Source-read log

No frozen-source files were read for this classification. The supplied records, their annotations and reconciliation supplied the warrants. No additional records, quotations or evidence were introduced.