Refresh batch 01, rerun: Agent-S, MemoryOS and Basic Memory

The operator commissioned this handoff on 2026-09-28 for execution in a new session. Its purpose is to rerun the first refresh batch under the current method. The first run of batch 01 (branch refresh-batch-01, commit a39ba9be, record kb/work/agentic-memory-refresh/batch-01-2026-09-28.md on that branch) published all three sets, but in the layout that preceded the directory-artifact output, and its friction led to fixes the method now carries: worktree library resolution, line ranges only in quote attributions, no Run identity section, no trailing whitespace in generated quotes, a member lookup table, and finalization rules for merged and rejected proposals. Its sets are superseded, not migrated; that branch is not merged. This rerun analyses the same three systems afresh, two at a time, each producing a retained analysis set and a fresh public review. No analysis of this rerun has been started.

Command for the next session

Paste this instruction into a fresh session at the Commonplace repository root:

Run the refresh batch 01 rerun described in
kb/work/agentic-memory-refresh/batch-01-rerun-handoff.md. Act as the batch
coordinator: create the batch worktree, check the preconditions inside it,
run the three analyses two at a time, each with one fresh source-only
coordinator and its fresh memory specialist; verify each published set;
run the bounded downstream checks; update the three inventory rows; write
the batch record; and commit the batch on its branch. Follow the fixed
inputs and boundaries in that file.

Batch worktree

The batch runs in its own git worktree so that other sessions' edits in the main checkout cannot block or contaminate it. From the main checkout:

git worktree add -b refresh-batch-01-rerun ../commonplace-refresh-batch-01-rerun HEAD

Every analysis, verification and write happens inside ../commonplace-refresh-batch-01-rerun. Its base commit is every run's inputs-commit. The installed commonplace-* commands execute the main checkout's src/commonplace/, which must stay at the base commit for the whole batch; publication refuses otherwise, so if it moves, stop and report. Commands run inside the worktree use the worktree's kb/ as the library without any override.

Source checkouts stay under the main checkout's related-systems/ and are passed to coordinators by absolute path. Run state under kb/reports/state/ is ignored and lives in the worktree.

Preconditions

Inside the batch worktree, confirm the following and stop with a report if any fails:

  • kb/reports/types/agentic-system-analysis-set.md exists, and skill step 7 writes output/overview.md, output/runtime.md, output/memory.md, output/epistemic.md and output/ARTIFACT.yaml.
  • commonplace-quote --help shows --selections.
  • Each source checkout below has the origin named below.
  • The worktree is clean, and the main checkout's src/commonplace/ has no difference from the worktree's base commit.

At startup record HEAD and the worker models.

Fixed inputs

System Repository Checkout Legacy artifact Expected class
Agent-S https://github.com/simular-ai/Agent-S <main checkout>/related-systems/simular-ai--Agent-S kb/agent-memory-systems/reviews/Agent-S.md (legacy revision 73ea1722…) agent harness with experience memory
MemoryOS https://github.com/BAI-LAB/MemoryOS <main checkout>/related-systems/BAI-LAB--MemoryOS kb/agent-memory-systems/reviews/MemoryOS.md (legacy revision 1d717060…) memory/knowledge/context-engineering system
Basic Memory https://github.com/basicmachines-co/basic-memory <main checkout>/related-systems/basicmachines-co--basic-memory kb/agent-memory-systems/reviews/basic-memory.md (legacy revision fc2ee070…) memory/knowledge/context-engineering system

This is a refresh: each coordinator fetches the current default branch of its origin, resolves it to a full commit, and freezes that commit as the run's source. Record it in the batch record. The legacy revisions are inventory metadata, not the analysis boundary.

The functional scope is fixed with the source: each analysis covers the whole shipped system at that commit, boundary-kind: whole-system, naming every excluded subsystem with the conclusion its exclusion prevents. A coordinator that finds the whole system infeasible stops and reports; the batch coordinator decides whether to accept subsystem-only and records the decision.

Public destinations are kb/agentic-systems/reviews/agent-s.md, kb/agentic-systems/reviews/memoryos.md and kb/agentic-systems/reviews/basic-memory.md. None exists on main; archived reviews under reviews-archive/ and the first run's sets on the refresh-batch-01 branch are not incumbents and are not to be read. Allocate fresh run IDs using the execution date.

Execution and isolation

Follow the current analysis skill and its contracts. Each analysis is one fresh coordinator plus its fresh memory specialist. Create each coordinator with fresh context (for the collaboration tool, fork_turns="none"). Supply only its source identity, absolute checkout path, public destination, fresh run ownership, the whole-system scope rule, and the current method instructions; supply repository doctrine explicitly if the runtime does not load it. Do not give any worker this document, the legacy review, the archived reviews, the first run's branch or record, or another system's run.

Run two systems at a time: two coordinators with their two specialists fit four worker slots. Start Agent-S and MemoryOS; start Basic Memory when a slot pair frees. If capacity is smaller, run sequentially and say so. Use completion events and owned output paths, not agent-status listings. Prior exposure to earlier analyses follows the skill's failure rule.

Do not edit governing files, the producer or the validators during the batch. Record a discovered defect with its run and let the run continue if it can complete under the contracts.

Defer every write of your own outside the run directories until all three runs have published: inventory, batch record, cache directory.

Acceptance

After each run, run the handoff command and commonplace-validate kb/reports/state/agentic-system-analysis/<run-id>/output --full, which validates the whole set including cross-member references, the profile's record references and the finalized memory member. Then:

  • Record per run: the resolved source commit; the boundary and its exclusions; body words and records per kind for each member; each failed validation or prepare with its first diagnostic; specialist correction turns; and every question, gap or invented value a worker produced, with the instruction text it was working from.
  • Downstream: run the matrix builder, table renderer and statistics script with exactly the three new reviews as --review arguments, into a fresh directory under kb/reports/cache/agentic-memory-refresh/, and confirm each exits 0.
  • Inventory: update only the three rows: status, new_run, new_review, and new_result pointing at the retained manifest kb/reports/retained/agentic-system-analysis/<run-id>/ARTIFACT.yaml, the file the review pins.

Write batch-01-rerun-<date>.md in this workshop, linked from its README, and answer: did every set validate and publish first time after preparation; which of the first run's friction points recurred and which are new; and the failed-validation, failed-prepare and correction counts per run beside the first run's (Agent-S 1 failed prepare and 2 corrections, MemoryOS 1 and 1, Basic Memory 1 failed prepare, 1 failed memory validation and 3 corrections). Trace hashes, per-turn telemetry and worker minutes are not required this time; report elapsed wall time for the batch.

Commit the batch on refresh-batch-01-rerun, staging by explicit path: three reviews, three retained sets, three inventory rows, the batch record and its README link. Do not merge, push or remove the worktree; the operator does.

This commission covers three analyses, their publication, the bounded downstream checks, three inventory rows, the batch record and the branch commit. It does not touch the archive or batch 01's branch, change the method, merge or delete the first run's branch, or commit anywhere else. Finish with the answers, the branch and commit hash, or concrete blockers.