Agent Skills for Context Engineering memory profile

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

Comparison rationale

This profile classifies the accepted records, not a deployed memory service. All ten inventories remain partial because the retained account leaves material shipped paths and consumer/admission branches unresolved. Known units describe only their named parts. The common unresolved breadth units preserve the memory report's included inventory; their citations anchor its scope limits rather than invent evidence about uninspected implementations.

The canonical novelty comparator is EPI-RTE-novelty-check, replacing the memory declaration through reconciliation. The canonical readiness predicate is EPI-RTE-run-readiness, replacing the workflow declaration. Their retained inputs establish ranking, validation and bounded enforcement, not semantic acceptance. Registry/ledger objects remain distinct from admission and from their proposal inputs. No superseded record supplies a profile finding.

Source review's claimed-versus-afforded disagreement remains unresolved. The profile therefore does not select either implementation characterization to establish human-controlled admission. EPI-RTE-source-review supplies the canonical protocol reference, not an authenticated manual-write witness. Reviewer strings and generic callers also leave initial content control unresolved.

The distributed numerical form of MEM-OBJ-vector-store is not a model-weight update. Graph storage and process memory coexist; Git-backed JSONL contributes repository and file substrates. CAS hashing classifies access metadata, not opaque body encoding. Numerical latent doctrine remains claimed, while actual payload provenance and persistence remain unresolved in MEM-CLM-context-protocols.

The durable trace-fed prompt chain supports trace-learning yes independently of demonstrated learning or exercised self-improvement. RT-RTE-prompt-optimization and MEM-OBJ-final-prompt establish automatic replacement, later capture supply and cross-invocation artifact consumption. MEM-OBJ-reasoning-trace preserves the difference between live input and truncated display. Activity is consumed as linked steps and tool records; mock outputs do not become verified environment outcomes. Generated skill files do not establish installed-host operation. No no finding is manufactured from the lack of observed benefit.

Requested retained-material consumption is wired for registry comparison, scaffold construction and separate final-prompt-to-skill generation. Plan reconstruction is a separate afforded pull unit bounded to its save/load demonstration, not deployed agent continuation. MEM-RTE-memory-access, MEM-OBJ-artifact-body and the observation/lazy-loading operations of MEM-RTE-context-selection establish access or reconstruction capabilities but leave later enclosing consumers unresolved; their direction units therefore have no controlled findings. Exact paths and IDs do not close those missing chains. Automatic successor-capture instruction supply is push, with a coarse whole-prompt selector. These are different operations even inside the same optimizer. The declared mirrored context path has its own claimed coarse push finding. Neither catalog IDs nor names on requested observation files establish identifier-based push. MEM-RTE-context-selection provides storage and transformation affordances without proving a universal successor delivery hook.

Projection and regex extraction consolidate existing content; prompt/property replacement evolves it; cleanup/FIFO forget it. MEM-ABS-helper-consolidation establishes absence only in the empty integrated method. It does not erase other supported transforms. Export invalidation remains a declared correction protocol in MEM-OBJ-public-projection, not automatic propagated withdrawal. Model generation alone establishes neither semantic synthesis nor accepted new knowledge.

Source-read log: no source files were read. Classification used only the supplied members, reconciliation, types and record contract; no new quotes, records or evidence were added. No target execution or provider call was performed.