Analyse agent memory

Type: kb/types/instruction.md

Establish the system's memory mechanisms and their supported comparison classifications for integration into the main agentic-system analysis.

Commission and boundary

Run as a fresh specialist under analyse-agentic-system. Require the parent run ID, frozen memory-input.md, report destination, and permitted source access. The input supplies the subject, source register with full revision or capture digest and access root, relevant canonical records, requested memory scope and depth, exclusions, and any specific question. Its records are provisional findings to check against sources, not accepted conclusions.

Write only the commissioned memory-report.md using kb/reports/types/agent-memory-analysis-report.md. Read that contract and the Memory comparison fields and Status fields sections of kb/types/agentic-system-analysis-result.md. Do not load the legacy review type, prior system reviews, surveys, matrix outputs, or style exemplars. The parent owns canonical IDs, integration, publication and completion. Do not publish, modify the parent's input/result, delegate, or stage and commit.

Hash the input and this skill before analysis. Report those identities and your actual model identity; state unknown if the runtime does not expose it. Verify input and method hashes again before returning. Changed input requires a new handoff from the parent, not silent reconciliation in the worker.

Inspect and explain

Use the frozen primary sources. For Git, inspect commit-addressed blobs through git --no-replace-objects -C <source-root> show <full-commit>:<path> and scoped grep or ls-tree; the worktree and current HEAD are not evidence. For captures, verify the supplied digest before reading. Select paths and ranges before reading; truncated output supplies no evidence until the needed range is delivered in a bounded read. A missing source or needed scope expansion returns a blocked report with the conclusion it prevents.

Follow the old memory review's analytical progression: core mechanisms, operative artifacts, write side, read-back, then a curiosity pass. Keep each finding source-native before giving a Commonplace classification. Retain the minimum verbatim supporting code or prose for each load-bearing finding as a blockquote ending with > --- `commit-relative/path` @ `full-commit` or a matching commit-pinned GitHub blob link (the frozen source identity for a capture). Do not include display line numbers or invented ellipses. Separate discontiguous passages. Publication searches the full frozen source for that text; checking a file and line range alone is insufficient. Explain the passage's support for the finding, and let the parent retain it once on the canonical record. A thin memory boundary warrants short sections with explicit limits.

• Core mechanisms: explain what retained material can change in later work. Account for context volume and selection complexity, provenance and trust controls, and human editing/adoption surfaces where material.
• Artifacts: distinguish raw traces from derived memory, content from access metadata, and opaque payloads from their readable display summaries. Record storage, representational form, derivation, and authority at the actual consumer. Reuse canonical IDs with a short description; propose missing records as MEM-OBJ-1, MEM-RTE-1, and analogous local IDs.
• Write side: identify producer, input, trigger, persistence, rejection, maintenance and withdrawal. Separate manual authoring, automatic acquisition, and automatic operations over already retained material. Examine every trace-fed transformation, including compaction, for a later consuming route. For each derived behavior-shaping artifact, state whether it retains the reason for what it prescribes and whether any later route reads that reason; the parent records this on the theory route, where a rationale a later route reads can guide diagnosis.
• Read-back: trace retained material through selection and delivery to a named later consumer. Separate availability, delivery, activation and demonstrated benefit. For pull, identify the requesting consumer role and supported interface; an API with an unspecified hypothetical caller is only a storage capability. A documented external consumer role may establish an afforded route without deployed wiring. Push requires an automatic selector; name its trigger, inputs, selected parts, budget and consumption channel.
• Curiosity: challenge strong source claims, misleading labels and partial ontology mappings. Keep current Commonplace recommendations outside this report. No comparison to other systems is needed.

Classify and hand back

Fill all fourteen comparison axes in the report's memory-comparison, using the main-result contract's assessments, bases and controlled values. You own the proposed classifications as well as their supporting analysis. Give each known value its supporting records and inference. A complete set covers all scoped alternatives; use explicit uncertainty when an included branch remains opaque or uninspected. A Session identifier alone supplies no task horizon. Distinguish missing evidence from a negative finding. Use local proposal IDs where the parent has not yet registered a discovered object or route. Write every identifier in full, including lists: MEM-OBJ-1, MEM-OBJ-2. Never use abbreviations such as MEM-OBJ-1/O2 or ranges. The parent maps exact tokens.

Use curation terms consistently: consolidate reduces retained content without new claims; dedup merges near duplicates; evolve revises an existing entry; synthesize creates a claim absent from the inputs; invalidate withdraws current reliance while retaining history; decay forgets or downweights; promote raises tier or salience. Index rebuilds and acquisition alone do not establish these operations. Trace-learning scope, timing and form must cover the same qualifying routes.

Record corrections, proposed records, unresolved questions and limitations inside the report. Questions that prevent integration set report-status: blocked; justified unknown classifications do not by themselves block a complete report. Validate the report with commonplace-validate --full <report-path> and correct structural errors. This is specialist analysis, not independent semantic clearance of the main result.

Return the report path, SHA-256, status and a short summary of integration issues. Progress and urgent scope/access requests may be sent separately, but every substantive finding or unresolved issue must be in the final report.