sqz memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the accepted records, not a new source inspection or an execution. All positive findings are implementation wiring or exposed capabilities. None claims observed model use, reliable reconstruction, improved capacity or causal benefit. Storage coverage is known only for the explicitly recorded inventory; unrestricted payload interpretation and incomplete admission/consumer chains keep the other inventories partial.
Canonical identity. Reconciliation supersedes the epistemic SessionState object with MEM-OBJ-session. All session-object classifications therefore use that canonical record. The narrower import and resume functions remain valid alongside MEM-RTE-session-continuity; no further supersession affects a value. Cache originals, compressed representations and metadata remain distinct rather than being replaced by the broad context-text container.
Memory versus current context. SQLite retention does not establish model-window membership. MEM-OBJ-known-files and MEM-BAP-reference-policy support routing, eligibility validation and enforcement at code consumers, not truthful residency or model compliance. Reference invalidation leaves payloads recoverable. Pins promote residency eligibility but do not inject the pinned body. MEM-RTE-cache-maintenance supports actual deletion separately from compact's hypothetical host evictions. The latter supplies no decay finding about a host window.
Trace-fed retention. RT-RTE-mcp-proxy, RT-RTE-cache-recovery and MEM-OBJ-cache-compressed jointly establish upstream tool results entering automatic compression and durable cache admission, followed by later compressed-baseline or recovery/context consumption. Under the profile's trace-learning definition, a retained continuation representation qualifies even when deterministic and even without a novel claim or measured improvement. That supports a local wired "yes" and tool-traces, not a conclusion about all cached inputs. Retained tool baselines have learning force at the durable-update consumer; returned text has knowledge force at the context consumer. The two authorities are separate paths.
The initial producer of full sessions remains unresolved. MEM-OBJ-session records caller/import acquisition and invalid synthetic dashboard rows, not an automatic complete trace capture system. MEM-OBJ-snapshot distinguishes a stored checkpoint API from the actual CLI projection. Imported session turns do not alone establish a qualifying session-log learning route; printing a guide does not prove its separate durable retention. MEM-OBJ-operation-metadata likewise does not connect raw regret logs or exposed recommendations to an automatically updated production policy. These limits prohibit complete learning/source coverage, but do not erase the wired tool-result witness. No accepted evidenced-absence declaration warrants a local trace-learning "no".
Curation and semantic limits. Projection of retained sessions is consolidation, including when raw sessions survive. CLI guide projection has wired support; the separate budgeted checkpoint builder and its SQLite persistence are afforded capabilities. Their storage, curation and lineage findings occupy separate units so that extracted classifications retain that evidence-strength distinction. Direct CLI guide and cache/access-structure derivation support wired other-compiled lineage; budgeted checkpoint derivation supports other-compiled only at the afforded basis. Same-ID upsert is evolution. Reference-residency markers invalidate reliance on an access premise, not source truth. Exact-repeat substitution, index upsert and cache collision checks do not themselves establish duplicate removal or merging. Meaning-changing compression has unresolved per-input semantics; no synthesize finding follows from a possible faithfulness defect. These distinctions follow MEM-OBJ-snapshot, MEM-RTE-session-continuity, MEM-OBJ-cache-compressed and MEM-RTE-cache-maintenance.
Requested versus automatic delivery. Recall, expansion and tool help are pull. Automatic baseline choice within a requested read/transform is still fulfillment of that request, not independent push. By contrast the installed compact SessionStart hook invokes resume without an explicit request for retained session knowledge. Its timestamp/latest-session and category/budget selector is coarse, not identifier-based targeting. Explicit session-ID resume is the separate pull variant. MEM-RTE-session-continuity supports the installed command; EPI-RTE-resume-guide limits the resulting influence to potential external adoption. Host delivery and actual model consumption remain unobserved.
Capabilities and unresolved alternatives. MEM-CMP-library-stores supports URL/search/tee storage and transformations at the afforded layer. Tee get is an actual file consumer, but ordinary automatic tee production is not established. The proxy's lazy full documentation and RT-CMP-wasm-engine memoization are process-local, unlike SQLite content; imported WASM turns/corrections lack an established behavioral consumer. None inherits native recovery guarantees. Opaque original bytes are not classified from a readable wrapper, and unknown caller humanity does not establish manual admission.
Source-read log. No frozen-source files were read for this profile. All classifications use the supplied accepted records and their retained passages; no new evidence, record or source identity was added.