Letta Code memory profile

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

Comparison rationale

Scope and coverage

The profile classifies retained memory rather than the entire runtime's storage or permissions. Local prose, symbolic access state and mutable policy remain separate operative parts. Attached local files do not disclose hosted payload encoding, provenance, maintenance or selection. Every axis is partial because positive local mechanisms coexist with unresolved included alternatives. “Known” units identify only their named scope and evidence layer; they do not upgrade another branch. The inventory and limits come from MEM-OBJ-hosted-memory, MEM-RTE-repository-sync, MEM-RTE-reflection-integration and MEM-RTE-memory-admission.

The mod-learning builder retains evaluation reports and can copy executable successor code, but the accepted memory account explicitly treats executable mods and package replacement as separate runtime change routes. They are not silently used to assign memory promotion, validation, parametric storage or trace learning here. The retained inference components likewise do not establish accumulated parameter updates. See EPI-RTE-mod-selection, RT-RTE-extension-load, RT-RTE-package-update and RT-CMP-turn-model.

Non-obvious mappings

Files and repo coexist for Git-backed memory: files supply bodies while the repository supplies retained revision/access structure. Raw session envelopes, summary prose and first-kept/progress metadata are not one representation or consumer. Trace-fed inputs have learning authority at their update consumer; resulting summaries supply knowledge at a later model, while retained preferences/procedures can supply instruction. This establishes delivery force, not compliance, factual acceptance or improved capacity. See RT-OBJ-local-transcript, MEM-OBJ-compaction-summary, MEM-BAP-memory-context and MEM-BAP-reflection-progress.

Compaction is wired consolidation and trace learning even though the summary's semantic fidelity is indeterminate. Its retained summary replaces active context and later reaches another invocation. This operational mapping neither claims synthesis nor resolves the epistemic assessment of the generated propositions. Reflection's automatic write/integration chain is wired; its requested contradiction resolution and duplicate filtering remain claimed curation findings. Root archival placement does not implement reduced reliance because root ARCHIVE.md remains automatically supplied. See MEM-RTE-transcript-compaction, EPI-RTE-summary-generation, EPI-RTE-reflection-guidance and MEM-CLM-archive-projection.

The manual finding is only afforded on the offered human arena selection path. Generic choice callers, force flags and policy-authorization environment values do not authenticate per-content human approval. Compaction and default reflection admission remain automatic despite optional human initiation. See MEM-RTE-arena-selection, MEM-RTE-memory-admission, EPI-RTE-memory-restore and EPI-RTE-summary-retention.

Read-back direction follows operations, not message role. Automatic core/history and skill advertisement are push; requested Memory/Skill/search returns are pull. A requested name/path does not make pull into identifier push. Identity matching is retained for the automatic conversation-history selection and progress-bounded reflection slice. Core layout and eligibility/budget filters remain coarse. Query-dependent local search is a ranking consumer of retained text, not inferred-lexical push. See RT-RTE-foreground-turn, MEM-RTE-core-projection, MEM-RTE-skill-loading, MEM-RTE-transcript-search and MEM-BAP-reflection-progress.

Session logs and tool traces name the original content actually supplied to summary/reflection updates. A compaction trigger, user-role summary wrapper or manifest checkpoint is not a new event-stream source. No recorded linked state/action/outcome training path establishes trajectories. Those omissions are not bounded negatives. See MEM-OBJ-reflection-payload and MEM-RTE-transcript-compaction.

Canonical records and remaining limits

Reconciliation supersedes the epistemic summary object with MEM-OBJ-compaction-summary, projection with MEM-RTE-core-projection, fetch with MEM-RTE-deferred-read, and arena choice route with MEM-RTE-arena-selection. Classifications cite those canonical records. These identity changes alter no controlled value and do not upgrade epistemic fidelity or actual chooser identity. The broader compaction/reflection chains coexist with their separate producer, retention, guidance and merge functions; those narrower records remain canonical.

Missing initial-content control prevents complete agency and lineage classifications. Uninspected replay/search-cache, normalization/multi-conversation selection, generic-worker repair, migration and feedback-consumer details prevent a complete transformation/consumer inventory. Opaque service state prevents external substrate, encoding, provenance and selection classifications. No evidenced-absence record supports a local trace-learning “no”; one known automatic route cannot make another unresolved route or its trace source inapplicable. See MEM-OBJ-core-memory, MEM-OBJ-agent-skills, MEM-RTE-reflection-integration, MEM-RTE-transcript-search and MEM-OBJ-hosted-memory.

Source-read log

No source files were read. All classifications use the supplied accepted records and their retained passages; no evidence, quotes or analyst records were added. Source execution, report verification, publication and coordinator-owned state remain outside this assignment.