Obsidian Second Brain review Crew logs and rule promotion

Local checks of shared Markdown logs, independent session reads, 24 concurrent writers and explicit lesson retention, including unread-history limits.

AgentIndex · Published 2026-11-05 · Updated 2026-11-05

Obsidian Second Brain stores agent knowledge in Markdown for people who need decisions, sources and context to survive individual sessions. Its Crew feature adds a shared job log that different agents can read. This provides an approachable starting point when a team wants inspectable files before adopting a separate memory service.

We evaluated Crew, not the full vault's semantic search or automatic rewriting. A plain folder retained all 24 jobs written by concurrent local processes. A submitted lesson remained pending until an explicit keep wrote it into the recipient's rules. We also found an important reading behavior: a limited unread read consumes older omitted records as well as those displayed.

Version and ten local checks

We pinned commit deff215377f0ecae68deed1074046f883425388e on November 5, 2026. The project configuration declares 0.20.0 and MIT licensing. Crew uses the Python standard library and documents Python 3.10 or later; we used 3.12.13. Pinned project description

The actual Crew command entry point ran with three filesystem resolvers redirected: its configuration directory, agent directories and rule paths. All destinations were temporary. We did not modify the real home directory, agent configuration or vault. Generated hooks were not loaded in a host, so successful configuration output does not establish model-visible context. The reproduction fixture records those boundaries and source hashes.

CheckObserved result
Repeated setupPreserved unrelated rules; no duplicated hooks
Same session reads unread jobs twiceJob absent from the second read
Another session of the same agentIndependently received the job
Newly submitted open lessonNo recipient rule change
Keep twicePromoted once; repeated decision unchanged
Drop after keepRejected without changing rules
24 concurrent local writesAll 24 job records retained
Malformed hook JSONSetup failed without rewriting existing rules

Together with record formatting and bounded-read checks, ten checks passed. All tasks were synthetic; no model or remote service was called.

A readable handoff record

Crew works without Obsidian, using daily Markdown logs and a lesson table in a plain folder. A job identifies its agent, area, completion summary and evidence path. That is easier to revisit than an unexplained assertion that a previous session already handled something.

We wrote one job and read it twice as recipient session a. The second read omitted it, while session b could still read it independently. This supports separate context consumption by different sessions. It does not establish automatic task allocation or prevention of duplicate work. Log and reader implementation

Write factual completion records with useful evidence locations. A path is a clue to inspect, not proof that the file exists or still represents the current project. Keep long conversation transcripts out of the handoff unless they are needed to understand a specific decision. The shared log preserves claims; it does not independently verify them.

Accepting a lesson changes its status

Our submitted lesson did not alter the recipient's rules. After keep L1, its text appeared in the temporary rule file. Repeating keep did not duplicate it, and an attempt to change the final decision to drop was rejected. Pending suggestions and accepted conventions therefore had observably different effects.

The human decision boundary depends on caller behavior. The CLI does not authenticate a human: a program with command access could invoke keep too. Host instructions should reserve that decision for the user. This is a workflow convention, not an identity or permission boundary. Crew usage and limitations

Good retained lessons are specific, supported and scoped. A temporary workaround from one task should not automatically become a permanent instruction for every future task. Periodically revisit accepted rules as dependencies and project conventions change.

Limited reads consume omitted history

For a new session, we limited the unread response to two records. It returned the newest two and reported older unread records that were not shown. A second read then said there were no new records. The implementation records the prior unread set in the cursor, including entries omitted from the response.

Repeated limited reads therefore are not pagination through the backlog. Use a dated digest or read the log itself when a handoff requires completeness. This matters for a new session joining a mature project: a concise response may control context size while omitting an important earlier decision. No new records does not mean every historical entry has been displayed.

Concurrency and host activation

Six worker threads launched independent CLI processes to write 24 jobs. We subsequently found all 24 in the local log. That establishes the result on this machine and filesystem. It does not establish distributed consistency. The documentation warns that file-sync copies can conflict and that Crew does not resolve such conflicts.

Setup reruns were stable in the temporary configuration. Malformed JSON caused refusal before existing rules were rewritten. Those are useful configuration properties, but actual hook events, trust settings and model-visible context need testing in the chosen client. Crew also does not create a remote connection or start a background loop. Integration requirements

Adoption decision

Try Crew in a plain folder if the immediate need is to preserve completed work and explicitly accepted lessons in readable files. Once that process is reliable, decide whether the broader Obsidian workflow is useful. Our checks do not evaluate semantic retrieval, automatic knowledge rewriting or complete multi-host activation.

For retrieval and graph-oriented memory, continue with the Cognee review. Before adopting external skills, see the SkillSpector static-scan review. If the shared decisions concern an AI video project, the Toonflow canvas review covers a complementary artifact workflow.