swarmrecall-dream（蜂群回忆梦）

SwarmRecall Dream

Agent dreaming — memory consolidation, deduplication, pruning, contradiction resolution, and session summarization. Run during idle periods to keep your memory store lean and accurate.

Auto-Registration

Before making any API calls, check for a SWARMRECALL_API_KEY environment variable:

1. If SWARMRECALL_API_KEY is set, use it for all requests as a Bearer token.
2. If SWARMRECALL_API_KEY is not set, self-register by calling:

   POST https://swarmrecall-api.onrender.com/api/v1/register
   Content-Type: application/json

   { "name": "<your-agent-name>" }

1. The response returns { "apiKey": "...", "claimToken": "..." }. Save the apiKey to the SWARMRECALL_API_KEY environment variable for all subsequent requests. Do NOT write the key to disk or any file without user consent.
2. Tell the user: "SwarmRecall is set up! To manage your agent's data, visit swarmrecall.ai/claim with code: <claimToken>"

Authentication

All API requests require:

Authorization: Bearer <SWARMRECALL_API_KEY>

API Base URL

https://swarmrecall-api.onrender.com (override with SWARMRECALL_API_URL if set)

All endpoints below are prefixed with /api/v1.

Privacy & Data Handling

• All data is sent to swarmrecall-api.onrender.com over HTTPS
• Dream operations only read and modify your agent's existing data — no new external data is collected
• Archived memories are soft-deleted (recoverable) — never hard-deleted
• Data is isolated per agent and owner — no cross-tenant access
• The SWARMRECALL_API_KEY should be stored as an environment variable only, not written to disk

How Dreaming Works

Dreaming uses a two-tier architecture:

Tier 1 (server-side): SwarmRecall runs deterministic operations that don't need intelligence — finding duplicate clusters, decaying importance, pruning low-importance memories, cleaning orphaned relations, and surfacing candidates for agent review.

Tier 2 (agent-driven): You (the agent) orchestrate the intelligent operations — reading candidate clusters, reasoning about what to merge/keep/archive, writing summaries, and resolving contradictions. You use the standard memory/knowledge/learnings endpoints to make changes.

Endpoints

Dream Cycle Management

Start a dream cycle

POST /api/v1/dream
{
  "operations": ["deduplicate", "decay_prune", "summarize_sessions"],  // optional, defaults to config
  "thresholds": {           // optional, override defaults for this cycle
    "similarityThreshold": 0.90,
    "decayAgeDays": 30,
    "decayFactor": 0.95,
    "pruneThreshold": 0.05,
    "batchSize": 500
  }
}

Returns the dream cycle with status: "running". Only one cycle can run per agent at a time (409 if already running).

List dream cycles

GET /api/v1/dream?limit=20&offset=0&status=completed

Get a dream cycle

GET /api/v1/dream/:id

Update a dream cycle (report results, mark complete/failed)

PATCH /api/v1/dream/:id
{
  "status": "completed",
  "results": {
    "deduplicate": { "clustersFound": 3, "memoriesMerged": 5, "memoriesArchived": 5 },
    "decay_prune": { "memoriesDecayed": 12, "memoriesPruned": 2 },
    "summarize_sessions": { "sessionsProcessed": 4, "summariesCreated": 4, "memoriesDecayed": 0 },
    "durationMs": 4500
  }
}

Dream Config

Get dream config

GET /api/v1/dream/config

Update dream config

PATCH /api/v1/dream/config
{
  "enabled": true,           // enable/disable auto-dreaming
  "intervalHours": 24,       // how often to auto-dream (1-168)
  "operations": ["deduplicate", "decay_prune", "summarize_sessions"],
  "thresholds": {
    "similarityThreshold": 0.90,
    "decayAgeDays": 30
  }
}

Tier 1 Execute (run server-side ops)

POST /api/v1/dream/execute
{
  "operations": ["decay_prune"]   // optional, defaults to Tier 1 ops in config
}

Runs server-side operations (decay, prune, orphan cleanup) immediately. Returns results.

Candidate Endpoints

These return pre-computed analysis for agent-driven (Tier 2) operations.

Duplicate memory clusters

GET /api/v1/dream/candidates/duplicates?limit=50

Response:

{
  "clusters": [
    {
      "anchor": { "id": "mem_1", "content": "User prefers dark mode", "importance": 0.8 },
      "members": [
        { "id": "mem_7", "content": "The user likes dark mode themes", "importance": 0.5, "similarity": 0.94 },
        { "id": "mem_12", "content": "User said they prefer dark mode", "importance": 0.6, "similarity": 0.92 }
      ]
    }
  ],
  "totalClusters": 3,
  "thresholdUsed": 0.90
}

Unsummarized sessions

GET /api/v1/dream/candidates/unsummarized-sessions?limit=20

Response:

{
  "sessions": [
    { "id": "sess_1", "memoryCount": 14, "startedAt": "2026-03-15T...", "endedAt": "2026-03-15T..." },
    { "id": "sess_2", "memoryCount": 8, "startedAt": "2026-03-18T...", "endedAt": "2026-03-18T..." }
  ],
  "totalSessions": 2
}

Duplicate entity pairs

GET /api/v1/dream/candidates/duplicate-entities?limit=50

Response:

{
  "pairs": [
    {
      "entity_a": { "id": "ent_1", "type": "Person", "name": "John Smith", "properties": {} },
      "entity_b": { "id": "ent_5", "type": "Person", "name": "J. Smith", "properties": {} },
      "similarity": 0.93
    }
  ],
  "totalPairs": 1,
  "thresholdUsed": 0.92
}

Stale memories

GET /api/v1/dream/candidates/stale?limit=100

Response:

{
  "memories": [
    { "id": "mem_20", "content": "...", "importance": 0.12, "createdAt": "2025-12-01T...", "ageDays": 121 }
  ],
  "totalStale": 15,
  "decayAgeDaysUsed": 30
}

Contradiction pairs

GET /api/v1/dream/candidates/contradictions?limit=50

Response:

{
  "pairs": [
    {
      "memory_a": { "id": "mem_3", "content": "User's timezone is PST", "createdAt": "2026-01-15T..." },
      "memory_b": { "id": "mem_45", "content": "User's timezone is EST", "createdAt": "2026-03-20T..." },
      "similarity": 0.91,
      "contentDivergence": 0.67
    }
  ],
  "totalPairs": 1
}

Unprocessed memories (for knowledge graph enrichment)

GET /api/v1/dream/candidates/unprocessed?limit=100

Returns memories where metadata.dreamProcessedAt is null or before the last dream cycle.

Behavior

When to dream

• Between sessions: After ending a session and before starting the next one. This is the natural idle period.
• On idle detection: If your framework detects the agent is idle (no user interaction for a configured period), trigger a dream.
• Periodically: If auto-dream is enabled in your config, the server runs Tier 1 ops automatically on your interval. You can layer Tier 2 on top.
• On explicit request: When the user says something like "clean up your memory" or "consolidate what you know."

Full dream flow

Follow this sequence for a complete dream cycle:

1. Start the cycle:

   POST /api/v1/dream
   { "operations": ["deduplicate", "summarize_sessions", "decay_prune", "resolve_contradictions"] }

1. Run Tier 1 server-side ops (decay, prune, orphan cleanup):

   POST /api/v1/dream/execute

1. Fetch and process duplicate clusters:

   GET /api/v1/dream/candidates/duplicates

For each cluster: - Read all members, reason about which content to keep - Update the anchor memory with merged content: PATCH /api/v1/memory/:anchorId - Archive each duplicate: PATCH /api/v1/memory/:duplicateId with { "archived": true }

1. Fetch and process unsummarized sessions:

   GET /api/v1/dream/candidates/unsummarized-sessions

For each session: - Read its memories: GET /api/v1/memory?sessionId=<id>&limit=100 - Reason about the session's key decisions, facts, and outcomes - Write the summary: POST /api/v1/memory with { "content": "<summary>", "category": "session_summary", "sessionId": "<id>", "importance": 0.7 }

1. Fetch and resolve contradictions:

   GET /api/v1/dream/candidates/contradictions

For each pair: - Read both memories and their context (timestamps, sessions) - Decide which is the current truth - Archive the stale one: PATCH /api/v1/memory/:staleId with { "archived": true } - Optionally update the current one with a note about what changed

1. Complete the cycle:

   PATCH /api/v1/dream/:cycleId
   {
     "status": "completed",
     "results": { ... counts of what you did ... }
   }

Tips

• Start small: For your first dream, just run decay_prune and deduplicate. Add more operations as you get comfortable.
• Check before you merge: Always read the full content of duplicate clusters before merging. High similarity doesn't always mean the memories are truly redundant — they might add different context.
• Protect important memories: Add the "pinned" tag to any memory that should never be pruned. Session summaries are automatically protected.
• Report results: Always update the dream cycle with results when done. This creates an audit trail and helps the dashboard show what happened.
• Don't dream too often: Once every 24 hours is a good default. More frequent dreaming wastes compute with diminishing returns.

Shared Pools

• Dream operations respect pool boundaries. Candidate endpoints only return data the agent has access to.
• Pool-scoped data (memories, entities, learnings with a poolId) is included in candidate analysis if the agent has readwrite access to the pool.
• When merging pool-scoped memories, ensure the merged result retains the poolId.