Case memo

The memo is where people teach a case how it should classify: filing rules, comments on results and corrections. It is a set of scoped entries with a revision history, available on any case. Classify jobs load the entries that apply to each row and record which memo version they used, so every result can be traced to the rules behind it.

Entries

Field	Values
type	instruction (a rule), feedback (a comment on a result), correction (created by a review; a label fix with its reason)
scope.kind	global; subtree with a category code (that category and its descendants); record with a record key; document with a documentId
status	proposed, active, retired
text	Up to 8,000 characters. Stored in blob storage, not in the database.
links	Optional classificationResults (jobId, rowId) and recordKeys the entry relates to

subtree and record scopes only make sense on a records case. Every change is a new revision; the memo version is the highest revision of the case, and the entries active at any version can be read back.

Lifecycle:

• An instruction starts as proposed and becomes active only when a human approves it. Nothing changes in classification until then.
• Feedback is stored active and is inert: it does not change classification. It exists to be turned into an instruction.
• A correction is created by a classify review and applies directly.
• Retiring an entry adds a revision with status retired; earlier versions stay readable.

API

Method	Path	Notes
GET	/cases/{caseId}/memo	Filters type, scope, status; version= reads the snapshot at that memo version. Returns memoVersion and items.
POST	/cases/{caseId}/memo/entries	Create an instruction or feedback entry. Returns 201 with id, status, revision, memoVersion.
POST	/cases/{caseId}/memo/entries/{entryId}/approve	Approve a proposed instruction. Human credentials only. 409 if the entry is not a proposed instruction.
POST	/cases/{caseId}/memo/entries/{entryId}/retire	Retire an entry. 409 if it is already retired.
GET	/cases/{caseId}/memo/entries/{entryId}/revisions	Every revision of an entry.

POST /cases/42/memo/entries
{
  "type": "instruction",
  "scope": {"kind": "subtree", "code": "7.31"},
  "text": "Fiction books are filed by genre when the description names one; otherwise use 7.31.99.",
  "links": {"classificationResults": [{"jobId": "55", "rowId": 17}]}
}

{"id": "m_88", "status": "proposed", "revision": 41, "memoVersion": 41}

Any organization member with access to the knowledge base can write entries. Agents act with a short-lived agent token that the server recognises. An agent can propose instructions and leave feedback, but approving returns 403. The X-DocuDevs-Actor header is ignored. See Agent actions.

How a classify job uses the memo

A job does not load the whole memo. For each row it loads the active entries that apply: global entries, subtree entries whose subtree contains a candidate category of the row, entries scoped to the row's own key, and entries scoped to the row's document. Active corrections scoped to a record or subtree are loaded like instructions; feedback never is. Order is global first, then by subtree depth.

A token budget caps the memo text per row (about 1,500 tokens). If it is exceeded, the more specific entries win and the row gets the memoTruncated flag. The job pins the memo version when it starts (config.memo.version pins an older one, use: "none" ignores the memo), and every result stores the memo version and the ids of the entries it used.

When a job classifies rows of one case against another case's knowledge base, the knowledge base's entries are loaded first, then the work case's global and document entries.

To try a proposed instruction before approving it, list its id in config.memo.include, run the job, and compare the results with a run that leaves it out. To measure the effect over a grouped holdout sample before approving, use the memo A/B evaluation.

When a category tree is activated, the activation report lists memo entries whose scope names a retired or missing category. DocuDevs does not remap them: they keep working on the old codes until you update or retire them.

Python SDK

entry = await client.add_memo_entry(
    case_id,
    "instruction",
    {"kind": "subtree", "code": "7.31"},
    "Fiction books are filed by genre when the description names one.",
)
await client.approve_memo_entry(case_id, entry["id"])

memo = await client.list_memo(case_id, status="active")
history = await client.list_memo_revisions(case_id, entry["id"])
await client.retire_memo_entry(case_id, entry["id"])

CLI

docudevs memo add 42 instruction "Fiction books are filed by genre." --scope subtree --code 7.31
docudevs memo list 42 --status proposed
docudevs memo approve 42 m_88
docudevs memo revisions 42 m_88
docudevs memo retire 42 m_88