ocas-elephas（奥卡斯埃法斯）

Elephas

Elephas is the system's long-term memory — the sole writer to Chronicle, the authoritative knowledge graph where entities, relationships, events, and inferences live permanently once confirmed. It ingests structured signals from every skill's journals, scores candidate facts for confidence, resolves identity across possible duplicates, and promotes what survives into durable Chronicle facts with full provenance. The Chronicle database initializes automatically on first use — no manual setup required.

When to use

• Query Chronicle for entities, relationships, events, or inferences
• Ingest new signals from skill journals
• Run consolidation passes on pending candidates
• Resolve entity identity (possible duplicates)
• Promote or reject candidates
• Check Chronicle health and pending queue

When not to use

• Social relationship queries — use Weave
• Web research — use Sift
• Person-focused OSINT — use Scout
• Direct user communication — use Dispatch

Responsibility boundary

Elephas owns Chronicle: the authoritative long-term knowledge graph for entities, places, concepts, things, and their relationships.

Only Elephas writes to Chronicle. All other skills are read-only consumers.

Elephas does not own the social graph (Weave), OSINT briefs (Scout), or web research (Sift).

Elephas and Mentor are parallel journal consumers. Elephas reads journals to extract entity knowledge. Mentor reads journals to evaluate skill performance. Neither blocks the other.

Storage layout

~/openclaw/db/ocas-elephas/
  chronicle.lbug      — Chronicle graph database (auto-created on first use)
  config.json         — consolidation and inference configuration
  staging/            — temporary files during ingestion passes
  intake/             — incoming signals from other skills
    {signal_id}.signal.json
    processed/        — moved here after ingestion

~/openclaw/journals/ocas-elephas/
  YYYY-MM-DD/
    {run_id}.json     — one Action Journal per consolidation or promotion run

Default config.json:

{
  "skill_id": "ocas-elephas",
  "skill_version": "2.3.0",
  "config_version": "1",
  "created_at": "",
  "updated_at": "",
  "consolidation": {
    "immediate_interval_minutes": 15,
    "deep_interval_hours": 24
  },
  "identity": {
    "auto_merge_threshold": 0.90,
    "flag_review_threshold": 0.70
  },
  "inference": {
    "enabled": true,
    "min_supporting_nodes": 3
  },
  "retention": {
    "days": 0
  }
}

Database rules

LadybugDB is an embedded single-file database. One READ_WRITE process at a time. Other skills open chronicle.lbug as READ_ONLY only — Elephas holds the READ_WRITE connection during active passes.

Surface lock errors immediately. Do not retry silently.

Auto-initialization

Every command that opens the database runs _ensure_init() first. No manual init needed on first use.

Read references/init_pattern.md for the _open_db implementation pattern. Full DDL is in references/schemas.md.

Commands

elephas.ingest.journals -- Ingest structured signals from skill journal files and signal intake directory. Read references/ingestion_pipeline.md. Auto-inits on first call. Writes Action Journal.

elephas.consolidate.immediate -- Immediate consolidation pass. Score candidate confidence, promote above threshold, flag possible matches. Writes Action Journal.

elephas.consolidate.deep -- Deep pass: full identity reconciliation, inference generation, graph cleanup. Writes Action Journal.

elephas.identity.resolve -- Attempt to resolve whether two Entity records refer to the same real-world entity. Read references/ingestion_pipeline.md → Deduplication. Never silently collapse records. Writes Action Journal.

elephas.identity.merge -- Merge two confirmed-same Entity records. Always reversible. Append to merge_history. Writes Action Journal.

elephas.candidates.list -- List pending candidates by confidence tier and age.

elephas.candidates.promote -- Manually promote a candidate to a confirmed Chronicle fact. Requires at least one supporting signal. Writes Action Journal.

elephas.candidates.reject -- Reject a candidate with stated reason. Writes Action Journal.

elephas.query -- Query Chronicle for entities, relationships, events, or inferences. Read references/schemas.md for node types and Cypher patterns. All queries are read-only. Returns only confirmed facts unless include_candidates=true specified.

elephas.init -- Diagnostic and repair command. Checks schema, creates missing tables, verifies indexes. Use when troubleshooting — the database initializes automatically on first use.

elephas.status -- Report Chronicle health.

CALL show_tables() RETURN *;
MATCH (e:Entity) RETURN count(e) AS entities;
MATCH (p:Place) RETURN count(p) AS places;
MATCH (c:Concept) RETURN count(c) AS concepts;
MATCH (s:Signal {status: 'active'}) RETURN count(s) AS pending_signals;
MATCH (c:Candidate {status: 'pending'}) RETURN count(c) AS pending_candidates;
MATCH ()-[r]->() RETURN count(r) AS relationships;
CALL show_warnings() RETURN *;

Also report: last consolidation timestamps, pending identity reviews, inference count, intake queue depth.

elephas.journal -- Write Action Journal for the current run. Read references/journal.md. Called at end of every consolidation, promotion, merge, or rejection run.

elephas.update -- Pull latest skill package from GitHub source. Preserves journals and data.

Run completion

After every Elephas command that modifies Chronicle or processes signals:

1. Process all files in ~/openclaw/db/ocas-elephas/intake/; move processed files to intake/processed/
2. Persist ingestion results, promotion decisions, and merge records
3. Log material decisions to decisions.jsonl (if data directory exists)
4. Write journal via elephas.journal

Memory lifecycle

Skill Journal / Signal Intake
  → Signal (immutable after creation)
    → Candidate (pending until confirmed, rejected, or merged)
      → Chronicle Fact (persists indefinitely)
      → Inference (separate, never overwrites facts)

Consolidation passes

Immediate (every 15 min) -- creates candidates, scores confidence, promotes high-confidence. Scheduled -- promotes remaining, deduplicates. Deep (every 24 hr) -- full identity reconciliation, inference generation, graph cleanup.

Identity resolution rules

States: distinct (default), possible_match, confirmed_same. Merges are always reversible.

Resolution precedence: exact identifier match → name+location with corroboration → behavioral pattern match.

Ambiguous cases preserve separation. Never silently collapse records.

Write authority

Only Elephas writes to Chronicle. Other skills open chronicle.lbug as READ_ONLY only.

Elephas does not write to any other skill's database.

OKRs

Universal OKRs from spec-ocas-journal.md apply. Elephas-specific:

skill_okrs:
  - name: promotion_precision
    metric: fraction of promoted candidates uncontradicted after 30 days
    direction: maximize
    target: 0.90
    evaluation_window: 30_runs
  - name: identity_merge_accuracy
    metric: fraction of auto-merges not subsequently reversed by human review
    direction: maximize
    target: 0.95
    evaluation_window: 30_runs
  - name: candidate_queue_age
    metric: median age of pending candidates in hours
    direction: minimize
    target: 24
    evaluation_window: 30_runs
  - name: ingestion_coverage
    metric: fraction of journal files ingested within one consolidation cycle
    direction: maximize
    target: 0.99
    evaluation_window: 30_runs

Optional skill cooperation

• All skills — ingest structured signals from skill journals and signal intake
• Weave — read-only cross-DB queries for social graph enrichment
• Mentor — Mentor reads Chronicle (read-only) for evaluation context

Journal outputs

Action Journal — every consolidation, promotion, merge, rejection, and ingestion run.

Initialization

On first invocation of any Elephas command, run elephas.init:

1. Create ~/openclaw/db/ocas-elephas/ and subdirectories (staging/, intake/, intake/processed/)
2. Write default config.json with ConfigBase fields if absent
3. Create ~/openclaw/journals/ocas-elephas/
4. Open database with _open_db() which auto-creates chronicle.lbug and runs DDL if needed
5. Register cron jobs elephas:ingest, elephas:deep, and elephas:update if not already present (check openclaw cron list first)
6. Log initialization as a DecisionRecord

Background tasks

Cron options: sessionTarget: isolated, lightContext: true, wakeMode: next-heartbeat.

Registration during elephas.init:

openclaw cron list
# If elephas:ingest absent:
openclaw cron add --name elephas:ingest --schedule "*/15 * * * *" --command "elephas.ingest.journals && elephas.consolidate.immediate" --sessionTarget isolated --lightContext true --wakeMode next-heartbeat --timezone America/Los_Angeles
# If elephas:deep absent:
openclaw cron add --name elephas:deep --schedule "0 4 * * *" --command "elephas.consolidate.deep" --sessionTarget isolated --lightContext true --wakeMode next-heartbeat --timezone America/Los_Angeles
# If elephas:update absent:
openclaw cron add --name elephas:update --schedule "0 0 * * *" --command "elephas.update" --sessionTarget isolated --lightContext true --timezone America/Los_Angeles

Self-update

elephas.update pulls the latest package from the source: URL in this file's frontmatter. Runs silently — no output unless the version changed or an error occurred.

1. Read source: from frontmatter → extract {owner}/{repo} from URL
2. Read local version from skill.json
3. Fetch remote version: gh api "repos/{owner}/{repo}/contents/skill.json" --jq '.content' | base64 -d | python3 -c "import sys,json;print(json.load(sys.stdin)['version'])"
4. If remote version equals local version → stop silently
5. Download and install:

   TMPDIR=$(mktemp -d)
   gh api "repos/{owner}/{repo}/tarball/main" > "$TMPDIR/archive.tar.gz"
   mkdir "$TMPDIR/extracted"
   tar xzf "$TMPDIR/archive.tar.gz" -C "$TMPDIR/extracted" --strip-components=1
   cp -R "$TMPDIR/extracted/"* ./
   rm -rf "$TMPDIR"

1. On failure → retry once. If second attempt fails, report the error and stop.
2. Output exactly: I updated Elephas from version {old} to {new}

Visibility

public

Support file map