MindBase memory profile

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

Comparison rationale

This profile classifies the accepted records, not a deployed corpus. All positive implementation findings are conditional capacities: none has an observed or causal basis. The hosted builder sequence remains afforded even though its MCP replacement tool is implemented. The same value may therefore appear with separate bases. See RT-RTE-plugin-build and RT-RTE-build-context.

Reconciliation proposes no supersessions. The content objects, mixed-object parts and overlapping routes retain their declared identities. In particular, MEM-OBJ-context and EPI-OBJ-wiki-pages have different corpus coverage; MEM-RTE-synthesis-cache spans generation and retention whereas EPI-RTE-topic-synthesis isolates generation/admission. Classification uses their local parts rather than treating one as a stronger replacement.

Filesystem JSON is a storage witness, not a representational form. Dense cache vectors are parametric representations and vector access structures, while their paths/hashes are symbolic. Graph rows are graph access state implemented in SQLite. Model weights are external machinery, not accumulated MindBase memory. MEM-OBJ-embeddings and MEM-OBJ-access-metadata do not establish that every mutation invalidates every cache.

Retained project README guidance is instruction at the proposal model. Retained knowledge is knowledge at the answering model. Text and vectors also have ranking force at retrieval code; retained edit-state metadata supplies validation and an enforced veto at the inspected health handlers. These forces follow actual consumers, not file ownership or a generic knowledge label. See RT-BAP-project-rules, RT-RTE-ask-answer, MEM-OBJ-embeddings and EPI-RTE-health-maintenance.

The trace-learning positive is narrow. RT-RTE-ask-answer wires recorded conversation history into an answer invocation and conditionally retains marked output for later question retrieval. On that history-fed branch, automatic production of durable later context qualifies without new knowledge, a continuation-summary label or measured improvement. History has learning authority at this production consumer; saved output has knowledge authority downstream. This does not turn all source-to-context synthesis into trace learning: MEM-OBJ-sources leaves contributed-content origin mixed or unknown. Nor does MEM-RTE-chat-retention establish automatic trace derivation merely by saving caller records. The partial inventory supports a system-union yes, never a complete-value claim or a system-wide no.

Curation is equally route-specific. Direct changes and cached report replacement support evolve. Build guidance supports requested consolidation only; semantic preservation and novel claims are unresolved. Topic generation is not classified as synthesize merely because it is named synthesis, and return-only merge/flag handlers do not establish deduplication or history-preserving invalidation. See MEM-RTE-tree-edit, EPI-RTE-synthesis-cache, MEM-OBJ-context, EPI-RTE-topic-synthesis and EPI-RTE-health-maintenance.

Question/search delivery fulfills requests and is pull even when code automatically selects candidates. The session hook instead triggers on startup, clear or compact and supplies selected-project core without a contemporaneous file request. Its project identity match and coarse whole-core delivery are separate selection operations. Lexical and vector ranking on requested retrieval is not a push signal. See RT-BAP-session-context, RT-RTE-ask-answer and MEM-RTE-mcp-retrieval.

Unresolved units preserve the operational-loop boundary. Missing records for filing, schema edits, classification guidance and other worker/compiler alternatives prevent complete authority, admission, lineage and trace-source inventories. Missing body/output pairs prevent stronger meaning-level curation claims. Layout compatibility, snapshot success and cross-writer invalidation remain local limits, not absence. Context size checks apply only to named branches; the internal wikilink append and direct tree edit prevent inferring a universal cap or snapshot guarantee. See MEM-RTE-cli-ingest, EPI-RTE-health-maintenance, MEM-RTE-tree-edit and RT-RTE-build-context.

Source-read log: none. This classification used only the supplied accepted members and their contracts; no source read was needed to resolve a record ambiguity.