Community And Dreaming

Community Memory

This page covers community summaries, community-aware recall, scheduled community detection, and operation logs.

Dreaming now has its own page: Dreaming.

Detection And Summaries

Community detection groups related graph nodes inside one authenticated profile. It uses Neo4j Graph Data Science Leiden detection, then persists bounded community summaries in the same profile scope.

Community summaries are retrieval aids. They are not facts, claims, or evidence. Run detection again when the graph changes enough that the old community assignments are stale.

Use MCP tools:

{"tool": "detect_community", "arguments": {}}

{"tool": "list_communities", "arguments": {"limit": 20}}

{"tool": "get_community_summary", "arguments": {"community_id": "community-1"}}

Use REST to read persisted summaries:

curl "http://127.0.0.1:8080/api/v1/communities?limit=20" \
  -H "Authorization: Bearer $DENSE_MEM_API_KEY"

Community detection requires write scope because it updates graph community assignments and stored summaries. Listing and reading communities require read scope.

Community-Aware Recall

Normal recall does not expand through communities unless the caller asks for it. Set use_communities=true when a query may benefit from related graph clusters:

{
  "query": "deployment and release process",
  "limit": 10,
  "use_communities": true
}

For REST:

curl "http://127.0.0.1:8080/api/v1/recall?query=deployment&use_communities=true" \
  -H "Authorization: Bearer $DENSE_MEM_API_KEY"

Community expansion only fills unused recall slots. Direct fact, claim, and fragment matches still rank first.

Scheduled Community Detection

Scheduled detection is controlled from the control portal Config tab. It is disabled by default.

Setting	Default	Meaning
COMMUNITY_DETECTION_ENABLED	false	Enables scheduled detection across profiles.
COMMUNITY_DETECTION_START_TIME_LOCAL	03:30	Local HH:MM start time.
COMMUNITY_DETECTION_MAX_CONCURRENCY	1	Concurrent profile runs, from 1 to 8.
COMMUNITY_DETECTION_JITTER_SECONDS	600	Stable per-profile run jitter, from 0 to 3600 seconds.
APP_TIMEZONE	Local	Shared scheduler timezone. Use an IANA name such as America/New_York for predictable production behavior.
AI_COMMUNITY_MAX_NODES	500000	Environment cap for a projected community graph. Larger graphs are skipped.

The scheduler checks once per minute, reserves one run per profile and local run date, and stores reservations in Postgres so multiple server instances do not duplicate the same community run.

Dreaming

Dreaming creates reviewable hypotheses from existing memory. It is separate from community detection and from active facts. See Dreaming for the full workflow, tools, schedule, and review behavior.

Operation Logs

Operation logs are database-backed application logs for the control portal. They are separate from Prometheus telemetry.

The control portal Logs tab can filter by severity and sort by timestamp or severity. The Config tab controls retention:

Setting	Default	Meaning
OPERATION_LOG_RETENTION_DAYS	30	Number of days to keep operation logs, from 1 to 365.

The server flushes operation logs in batches and prunes old rows hourly.

Troubleshooting

Problem	Check
No scheduled community runs	COMMUNITY_DETECTION_ENABLED=true, APP_TIMEZONE, Neo4j GDS plugin availability, and AI_COMMUNITY_MAX_NODES.
Community-aware recall has no extra results	Run detect_community first and pass use_communities=true; expansion only fills unused recall slots.
Logs tab is empty	Generate server activity, then refresh; logs are flushed every few seconds.