MemPalace memory profile

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

Comparison rationale

Inventory and evidence strength

The profile classifies accepted records, not source execution. Drawer text, access metadata, embeddings and locator lines are distinct operative parts: MEM-OBJ-drawer-text, MEM-OBJ-drawer-metadata, MEM-OBJ-drawer-vectors and MEM-OBJ-closet-lines. Diary rows specialize drawer storage; SQLite temporal and coordination objects are independent stores. The filesystem tunnel sidecar is not classified as a graph substrate merely because its encoding expresses links. Media descriptors classify references, not opaque file contents. See MEM-OBJ-diary-entry, MEM-OBJ-kg-triples, RT-OBJ-task-request, MEM-OBJ-tunnels and MEM-OBJ-media-asset.

Every axis is partial because its inventory retains unresolved included alternatives. Backend names do not establish native/remote retention semantics. Initial file and generic caller control are unresolved. Provider request retention, complete graph policy, every host wrapper, full recovery/replacement and all event/artifact consumers are incompletely established. These limits are carried by MEM-OBJ-drawer-vectors, RT-RTE-backend-registration, MEM-OBJ-write-log, MEM-RTE-memory-maintenance, MEM-CMP-closet-provider, RT-OBJ-patch-artifact and the route-specific citations in the units. Static embedding/provider weights are computation, not evidence of accumulated model-weight memory. No observation or causal support is asserted.

Admission, content and effects

An operator starting import or regeneration does not approve each generated item. Software capture and silent checkpoint writes are automatic. Interactive setup confirmation is manual; its --yes alternative is automatic. Generic foreground tool callers and initial daily summaries do not establish human control. Identity writing, restructuring approval and task approval are separately claimed policies rather than universally enforced admission gates. See RT-RTE-hermes-capture, MEM-RTE-closet-regeneration, RT-RTE-init-detection, MEM-RTE-diary-checkpoint and WF-RTE-task-create.

Knowledge means a retained part supplies context or reference at its consumer, not that its propositions were accepted as true. Identity supplies guidance; worker task instructions have a documented reload consumer, not an observed reload. Ranking is established by stored words, vectors and weighted metadata consumed by fixed scoring code. Duplicate checks separately give retained content validation and enforcing admission effects; ordinary primary-key idempotency alone is not curation deduplication. See RT-BAP-agent-context, RT-BAP-task-handoff, RT-RTE-palace-recall, EPI-RTE-duplicate-check and EPI-RTE-checkpoint-disposition.

Consolidation records reduced retained representations, not merely transient wake rendering. Checkpoint prefixes and generated locator lines reduce existing source material; extraction fidelity defects do not establish synthesis. Explicit replacement evolves entries. Temporal ending invalidates current eligibility while preserving history. The corrected successor-reuse branch does not guarantee visibility at the requested boundary, so neither enforcement nor evolution implies unconditional replacement coverage. See MEM-RTE-source-filing, MEM-RTE-diary-checkpoint, MEM-RTE-closet-regeneration and MEM-RTE-temporal-facts.

Later delivery and trace-fed updates

Query-driven retrieval and host-requested prefetch are pull. Independent startup context supply is push; its selector uses availability, wing, recency/weight and budget. Query embeddings and lexical hydration are therefore not push-signal findings. A task/correlation identifier on a requested reload is not identifier-based push. See RT-RTE-palace-recall, MEM-RTE-wake-context and RT-RTE-task-launch.

Automatic checkpoint prefixes qualify as trace learning because recorded conversational turns produce a durable continuation artifact for later contextual use. Session-fed extracted segments and locators likewise have later recall/ranking consumers. This classification requires neither learned weights nor improved capacity. Their inputs have learning authority at the update consumer, while the outputs supply knowledge or ranking downstream. Raw transcript retention alone does not establish that update. Traversal counters have recorded audit/navigation display but no established downstream ranking effect; the profile retains that uncertainty instead of assigning a local or system-wide no. See MEM-RTE-diary-checkpoint, MEM-RTE-source-filing, MEM-RTE-closet-regeneration, RT-RTE-hermes-capture and MEM-RTE-tunnel-traversal.

The Stop event triggers checkpoint work but does not become the checkpoint's original trace source: selected user messages are session logs. Mixed source groups retain their uncertainty; imported project text is not an event stream. No qualifying tool-trace, event-stream or trajectory value is established solely by a callback, source label or stored event envelope. See MEM-RTE-diary-checkpoint and MEM-RTE-source-filing.

Canonical resolution and source-read log

Reconciliation resolves the duplicated epistemic drawer object to MEM-OBJ-drawer-text, the duplicated epistemic triple object to MEM-OBJ-kg-triples, and the duplicated epistemic temporal route to MEM-RTE-temporal-facts. The profile uses those canonical destinations. The temporal amendment changes identity resolution, not evidence strength, and keeps the corrected successor-reuse limitation. Startup, duplicate disposition and drawer revision remain valid parts/overlaps, not replacement identities; their local findings are not upgraded by a stronger container witness.

Source reads: none. The supplied accepted records and their annotations establish the classifications and limitations; no new evidence, record declarations or reopened verification were needed.