prose-contracts/memory-fixer.prose.md

report_only_agents, kind, name, description, version

report_only_agents	kind	name	description	version
koby	pattern	memory-fixer	Auto-fix low-hanging fruit in the RA-H OS knowledge graph. No judgment calls — only deterministic Level 1 operations. Escalate anything that needs Kwame's input. Executes confirmed Kwame decisions to completion (state + updated_at).	2.2.0

---

Memory Fixer

Executable copy: the okyeame-memory-fixer cron job on kagentz (hermes cron list) holds its instruction set inline in ~/.hermes/cron/jobs.json (hermes cron edit <id> --prompt …; there is no --prompt-file, and ~/.hermes/cron/memory-fixer-prompt.md is a synced draft, not the live instruction). This file is the institutional record of the same contract; when the two diverge, the job prompt is what actually runs — diff it against this file before claiming a prompt change landed. ⚠️ Corrected 2026-09-26: the previous pointer (/root/.hermes/contracts/memory-fixer-v3.md) does not exist on kagentz — no /root access from this container — and was verified unreachable, not merely stale.

Purpose

Auto-fix low-hanging fruit in the graph. No judgment calls — only deterministic Level 1 operations. Escalate anything that needs Kwame's input. When Kwame replies to an escalation, execute the decision to completion (update state and timestamps), never leaving a node in review-pending forever.

Type: Write-only (Level 1 fixes only) Scope: RA-H OS knowledge graph (192.168.68.65) Schedule: Daily at 8 AM ET Escalation: Level 2+ to Kwame as task items

Key Design Decision

The updateNode tool's metadata field performs a restricted merge — the state key only accepts 'processed' or 'not_processed'. Additionally, new metadata keys cannot be added via the merge.

Solution: Use the description field to tag stale nodes with review actions, since description is a simple string overwritable via updateNode.

Tag Format: [REVIEW: action] original description text...

Where action is one of:

• archive — node is stale and should be archived
• refresh — node is stale and should be refreshed (infrastructure)
• keep — node has been confirmed as current
• merge — node is a duplicate candidate

Query for finding review-tagged nodes:

SELECT id, title, description
FROM nodes
WHERE description LIKE '[REVIEW:%';

Level 1 Auto-Fixes (No Kwame Decision Needed)

1. Missing type Auto-Classification

SELECT id, title,
  CASE
    WHEN title LIKE '%infrastructure%' OR title LIKE '%proxmox%' OR title LIKE '%setup%' THEN 'infrastructure'
    WHEN title LIKE '%skill%' OR title LIKE '%how to%' OR title LIKE '%guide%' THEN 'skill'
    WHEN title LIKE '%doc%' OR title LIKE '%template%' OR title LIKE '%brand%' THEN 'documentation'
    WHEN title LIKE 'WAL:%' OR title LIKE 'TASK:%' THEN 'note'
    WHEN title LIKE '%[LEARN]%' THEN 'documentation'
    ELSE 'note'
  END as auto_type
FROM nodes
WHERE json_extract(metadata, '$.type') IS NULL;

2. Missing namespace Auto-Population

SELECT id, title, json_extract(metadata, '$.tenant') as tenant
FROM nodes
WHERE json_extract(metadata, '$.namespace') IS NULL;

3. Staleness Review Tagging (refresh-suggested nodes only)

Using the type-based windows from the memory-monitor contract, tag nodes stale beyond their window. Only process a maximum of 10 nodes per run to avoid overwhelming Kwame. Prioritize infrastructure first, then dynamic, then ephemeral.

Archive-suggested nodes are NO LONGER tagged — they are archived outright (see Level 1 fix 4). Tagging with [REVIEW: refresh] applies only to living nodes (infrastructure, deployment, system, system-health, business, philosophy, research, learning, investigation, analysis, project, agent, registry, policy).

Exclusion Rules:

• Nodes with state = review_pending, deprecated, archived, or not_processed are NOT processed
• Nodes whose description already starts with [REVIEW: or [ARCHIVED] are NOT re-processed

SELECT id, title, json_extract(metadata, '$.type') as node_type,
  CAST(julianday('now') - julianday(updated_at) AS INTEGER) as days_stale,
  CASE
    WHEN json_extract(metadata, '$.type') IN ('infrastructure', 'deployment', 'system', 'system-health') THEN 'refresh'
    WHEN json_extract(metadata, '$.type') IN ('business', 'philosophy', 'research', 'learning', 'learn', 'investigation', 'analysis', 'project') THEN 'refresh'
    ELSE 'archive'
  END as suggested_action
FROM nodes
WHERE updated_at < datetime('now',
  CASE
    WHEN json_extract(metadata, '$.type') IN ('infrastructure', 'deployment', 'system', 'system-health') THEN '-14 days'
    WHEN json_extract(metadata, '$.type') IN ('skill', 'documentation', 'template', 'protocol-enforcement', 'prd', 'architecture') THEN '-90 days'
    WHEN json_extract(metadata, '$.type') IN ('note', 'wal', 'WAL', 'task', 'TASK', 'event') THEN '-30 days'
    WHEN json_extract(metadata, '$.type') IN ('business', 'philosophy', 'research', 'learning', 'learn', 'investigation', 'analysis', 'project') THEN '-120 days'
    WHEN json_extract(metadata, '$.type') IN ('deprecated-relay', 'audit', 'audit-report', 'incident', 'incident-report') THEN '-3650 days'
    ELSE '-45 days'
  END
)
AND json_extract(metadata, '$.state') NOT IN ('review_pending', 'deprecated', 'archived', 'not_processed')
AND (description IS NULL OR description NOT LIKE '[REVIEW:%')
ORDER BY days_stale ASC
LIMIT 10;

For each identified node, call updateNode(id, { description: "[REVIEW: action] " + originalDescription }).

4. Stale-Node Archiving (Level 1 — standing Kwame directive, 2026-09-11)

Kwame's standing directive: stale nodes CAN be archived by the fixer. No per-batch escalation, no [REVIEW: archive] tagging — archive them.

For every node whose suggested action is archive (i.e. its type is NOT one of the living types in fix 3), archive it in a single updateNode call:

updateNode(id, {
  "description": "[ARCHIVED] " + originalDescriptionWithoutReviewTag,
  "metadata": {"state": "archived"}
})

• state transitions DO work through updateNode (archived, and back to active). The former "state only accepts processed/not_processed, use SSH" claim was wrong — verified 2026-09-11 by archiving 7 nodes (#61, #373, #388, #465, #475, #526, #1476) over the bridge with updated_at auto-bumping. SSH to the bridge host is a fallback, not a requirement, and it is blocked from kagentz anyway.
• Pass description and metadata in the same call, and always keep the updates object nested: {"id": N, "updates": {…}}.
• Archiving is non-destructive: the node stays in the graph, marked state: archived + [ARCHIVED] prefix. Living nodes (refresh-suggested) are NEVER archived without a specific Kwame decision — they are the cluster/agent/business canon.

Archive candidates are identified by the fix 3 query's suggested_action = 'archive' branch (the ELSE 'archive' case: anything not an infrastructure/skill/documentation/strategic/audit type).

5. Duplicate-Node Detection (Level 1 — read-only, every run)

The graph's duplicate problem is rarely an agent mistyping a title: it is recurring writers creating a new node per run instead of updating one. This phase detects that class and reports it. It is read-only and never merges.

python3 /home/hermes/.hermes/scripts/memory_dup_detect.py --json

Read-only, ~15s over the whole graph, exit 0. That script is the source of truth for the clustering logic — do not re-implement it in the prompt or hand-count "duplicates" from titles.

Consume each items[] entry's verdict field; do not invent your own:

verdict	Meaning	Required action
WRITER-DEFECT (run_family: true)	ONE scheduled task writes a new node per run	Report the ids, the agents (the writer) and span_days. Never merge — each node is that run's audit record. If the family grew since the last report, say UNFIXED and name the writer.
SAFE-MERGE	Bodies identical	Still requires an explicit merge #A into #B decision from Kwame.
HUMAN-DECISION	Same subject, bodies differ	Propose connect (an edge), never merge.

• Title overlap alone is not duplication. Four distinct client workflows of one family (#357-#361) and two different machines' migrations (#1792/#1793) both score high on title tokens while their bodies sit 0.1-0.3 apart. Confirm against body similarity before calling anything a duplicate.
• Report clusters as candidates for Kwame's decision, never as established duplicates — a wrong auto-merge destroys distinct content irrecoverably.
• Per-run history nodes are kept deliberately. Bulk-merging a run family destroys the audit trail the family exists for.

Level 2 Escalations (Kwame Decision Required)

1. Refresh-suggested stale nodes flagged with [REVIEW: refresh] — refresh or keep? (Archive-suggested nodes are auto-archived under fix 4 and are not escalated.)
2. Duplicate Nodes — as detected by fix 5, by verdict, never by raw title overlap. WRITER-DEFECT is a writer fix (update one canonical node), not a merge decision; SAFE-MERGE and HUMAN-DECISION clusters are escalated for merge-or-connect.
3. Orphan Nodes >90 days old — Archive or connect?

Reporting Format

The fixer does not send anything. Under the single-egress model (2026-09-21) every report leaves the node through Mumuni's gate (comms_drop.py for the queue, comms_gate.py to release and read-back verify), so exit 0 means QUEUED, never delivered. A report body is written to a file and handed to the outbox helper:

🦅 Memory Fixer — [HH:MM UTC]

Level 1 fixes applied:
- Missing type: X nodes classified
- Missing namespace: Y nodes populated

Stale nodes needing review (max 10):
1. [Node #XXX] Title — X days stale, SUGGEST: refresh
2. [Node #YYY] Title — Y days stale, SUGGEST: archive
...

Duplicate clusters (candidates — Kwame decides; the fixer never merges unilaterally):
1. [WRITER-DEFECT] #AAA/#BBB/#CCC — writer <agent>, N nodes, span Nd (UNFIXED if it grew since the last report)
2. [HUMAN-DECISION] #DDD/#EEE — same subject, bodies differ, SUGGEST: connect
3. "none" when the scan returned no clusters

Orphans >90 days:
1. [Node #EEE] Title — X days stale, orphaned

Reply with:
- "archive #XXX, #YYY" to mark for archive
- "archive all" to archive all stale nodes listed
- "keep #XXX" to confirm a node is current
- "merge #AAA into #BBB" to merge duplicates
- "refresh #XXX" to mark as current

Execution on Next Run

The fixer reads Kwame's previous response and executes the decision to completion — it must not leave a node in review-pending forever. Tagging alone is NOT enough; each confirmed decision must also update state and updated_at so the node drops out of the stale window on the next run.

⚠️ Corrected 2026-09-11: updateNode DOES accept state changes — {"updates": {"description": …, "metadata": {"state": "archived"}}} works over the bridge, and updated_at bumps automatically. The old "use direct SSH + SQLite for state transitions" instruction was based on a wrong assumption; SSH is a fallback only (and is blocked from kagentz). Use one updateNode call for both the tag and the state.

ssh root@192.168.68.65 "sqlite3 /root/.local/share/RA-H/db/rah.sqlite \"UPDATE nodes SET metadata = json_set(metadata, '$.state', '<state>'), updated_at = datetime('now') WHERE id = <id>;\""

Use updateNode only for description/source/title/link edits.

Decision → completed action mapping:

Kwame reply	Description change	State	updated_at
archive #XXX	replace [REVIEW: archive] → [ARCHIVED] prefix	archived	bumped to now
keep #XXX / refresh #XXX	clear the [REVIEW: …] tag entirely	active	bumped to now
merge #AAA into #BBB	set [REVIEW: merge_into #BBB] on #AAA, then follow manual merge workflow	handled manually	bumped to now
archive all	apply the archive row to every node listed in the prior report	archived	bumped to now

Why updated_at must be bumped (critical): the Level-1 staleness query keys off updated_at < now - window. If the fixer clears the tag but leaves a stale updated_at, the node is immediately re-flagged on the very next run and the cycle repeats forever. Bumping updated_at to now pushes the node back to the front of the window.

Exclusion after action: once an action is applied, the node's description no longer starts with [REVIEW: (archive → [ARCHIVED], refresh/keep → original text), so it is not re-processed.

After all actions are applied, verify with:

SELECT id, json_extract(metadata, '$.state') FROM nodes WHERE description LIKE '[REVIEW:%';

The result must be 0 rows when all decisions are executed. Report what was done.

Checks

• State integrity: archived nodes have state: archived + [ARCHIVED] prefix; kept nodes are state: active without a [REVIEW:] tag.
• Auto-archive applied: no node should ever be left tagged [REVIEW: archive] — that tag is retired. Any [REVIEW: archive] found means fix 4 was skipped; archive it and report.
• No review-pending forever: after executing Kwame's decisions, [REVIEW:% node count must be 0.
• Duplicate scan ran: every report carries the fix 5 block (none when there were no clusters). A report with no duplicate section means phase 5 was skipped — a silently skipped detection phase is the failure this phase exists to prevent.
• Timestamps: every executed decision (and every auto-archive) bumps updated_at, so the node exits the stale window on the next run.

Logging

Every Level 1 fix logged to ~/.hermes/logs/memory-fixer/YYYY-MM-DD.md Every Level 2 escalation logged and delivered to Kwame.