Cludebot memory profile

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

Comparison rationale

This profile classifies accepted records, not source files or observed runs. All positive mechanisms are implementation-grounded unless their own finding says afforded. All ten axes remain partial: inspected witnesses do not enumerate every included portable, client, initial-admission or later-consumer alternative. Inference weights are not accumulated memory merely because retained text passes through an inference model. Dense retained embeddings do have distributed-parametric form; their relational/vector storage is a separate classification (MEM-OBJ-retrieval-state).

Canonical references resolve the supplied reconciliation's supersessions: stored content uses MEM-OBJ-memories; main compaction uses MEM-RTE-dream-compaction; contradiction handling uses MEM-RTE-belief-revision; hosted and local queue processing use MEM-RTE-hosted-dream and MEM-RTE-local-dream. These mappings affect substrate, form, lineage, control, curation and trace-route classification without upgrading either analyst's evidence. The main cycle remains a valid container. Action measurement/refinement and shared-response/procedural consumption are overlapping but distinct routes, not duplicates.

Pull and push refer to different operations in a chain. A requested SDK return is pull. Automatic history/memory selection supplies the chat model without that model requesting individual records. Independent procedural supplementation similarly pushes selected rules into a requested generation; it is not an unsolicited user-facing background response (RT-RTE-chat-turn, MEM-RTE-response-memory). Snapshot generation is wired, but its later host supply remains only afforded (MEM-RTE-prompt-sync). No identifier signal follows just from an ID printed on returned data.

Curation classification follows transformation, not epistemic success. Main compaction reduces retained summaries even though fidelity and its hash pointer are unverified. Contradiction handling demonstrably downweights a source but does not implement a content-directed truth veto. Generated rules, journals and resolutions without inspected output instances do not establish synthesis of a new claim. Write duplicate checks and index construction alone do not establish deduplication (MEM-RTE-dream-compaction, MEM-RTE-belief-revision, RT-RTE-memory-entry).

Retained experience has learning authority at the automatic artifact-update consumer, while produced rules have instruction authority at the later response model. This supports wired trace learning, not demonstrated capacity improvement or exercised self-improvement. Conversational exchanges establish session-log input; social action/outcome records establish event-stream input. Mixed seeds and incompletely traced input origins prevent complete source coverage (EPI-RTE-action-learning, MEM-RTE-journal-continuity, MEM-BAP-procedural-guidance). SQLite queue completion lacks the recorded later consumer required for the same positive classification; it supplies neither an absence nor an inapplicability warrant.

Stored access controls have enforcing force on inspected guarded reads; a redelegation key is checked against retained encrypted content, giving that content a validation role. These are not factual validation or all-lane erasure. Portable revocation retains records and delegates filtering to applications, so universal enforcement is unresolved (MEM-RTE-memory-withdrawal, EPI-RTE-pack-verification). Generic caller identity supplies no explicit human content-decision witness for manual agency. Model-written task findings instead have automatic admission (WF-RTE-store-finding).

The reconciliation's pre-loop dashboard-recovery attribution conflict does not determine any memory value: this profile uses the distinct finding-write record, not an asserted shared recovery guarantee. No source reads were made; all warrants and limits come from the supplied members and their records. No new records, quotations, execution evidence or verification judgments are introduced.