pi-self-learning memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile classifies extension-maintained memory, not the entire host or completion provider. The retained-object inventory is the daily journal, canonical text and ranking parts, core and long-term views, monthly summaries and process-local notes. These records distinguish raw branch input from derived memory and distinguish canonical content from its lossy displays (MEM-OBJ-branch-messages, MEM-OBJ-daily, MEM-OBJ-index-text, MEM-OBJ-index-ranking, MEM-OBJ-core, MEM-OBJ-long-term-memory, MEM-OBJ-monthly, RT-OBJ-runtime-notes). Optional Git attempts support a conditional repository substrate, not an observed backup or read-back from commit history (RT-RTE-task-reflection).
Reconciliation supersedes the epistemic conversation-signals object with the memory branch-messages object, the epistemic monthly-text object with the memory monthly object, and the epistemic full-history claim with the memory complete-history claim. Citations use those canonical memory records. The parsed-reflection and aggregate-index containers remain valid, but classification uses their finer durable parts where consumers differ. These supersessions change citation identity, not evidence strength. The unresolved registration-count conflict does not change any classified memory route; this profile asserts no command count (MEM-CLM-complete-history, RT-RTE-runtime-controls).
Knowledge, instruction and learning describe different consumers. Historical prose is supplied to a foreground model; prevention advice can direct its conduct; retained experience also feeds further durable updates. None establishes truth, adherence or improved capacity. Static consultation instructions are excluded as shipped policy, and strict-mode wording supplies no retained-memory enforcement finding (RT-BAP-memory-policy, EPI-RTE-watchout-conversion, RT-RTE-task-reflection). Symbolic kind and ranking state act earlier at code consumers, so they retain routing/ranking findings independently of downstream prose authority (MEM-OBJ-index-text, MEM-OBJ-index-ranking).
Extraction, summarization, prefix retention, normalized-key merging, age downweighting and core promotion support curation findings. They do not resolve whether model text introduces new propositions or preserves every input claim. In particular, an Avoid prefix is an implemented policy conversion, not a semantic check; redistill's preservation request does not establish successful preservation. That unresolved semantic branch keeps curation partial without suppressing implemented operations (EPI-RTE-watchout-conversion, EPI-RTE-core-ranking, RT-RTE-month-summary, RT-RTE-core-redistill).
Requested acquisition and unsolicited model-input delivery can occur in one chain. A requested month supplies an identity selector that actually filters daily filenames before the summary model receives their content. Recency, rank, include flags and prefix budgets remain coarse signals; printing a file path or assigning rewrite IDs does not itself establish identifier-targeted push. Foreground consultation is an afforded pull route rather than another wired retrieval tool (RT-RTE-month-summary, RT-RTE-core-redistill, RT-RTE-context-injection, RT-BAP-memory-policy).
Trace learning is supported by automatic admission into durable reflection artifacts and later consumers, including when a human starts reflection. Original inputs include conversational turns and tool feedback. Later summaries and rewrites can inherit those sources through retained reflections, but legacy/imported inputs have unknown origin and admission control. This prevents complete lineage, write-agency and trace-source coverage. No model-weight learning or observed improvement is inferred (MEM-OBJ-branch-messages, MEM-OBJ-daily, MEM-OBJ-index-text, EPI-RTE-index-migration, RT-RTE-task-reflection, RT-CMP-reflection-model).
Source-read log: no frozen-source files were read for this mapping. All classifications use the supplied accepted records and their annotations; no records, quotations or new evidence were added.