Knowledge Assimilation

Universal Knowledge Assimilation Engine

Multi-source content discovery → classify → ingest → evolve pipeline. Extends the pull-based agent-utilities-evolution skill with a push-based pathway where any high-signal content from ANY source automatically triggers the self-evolution loop.

Architecture

graph TD
    A["Incoming Knowledge"] --> B["UniversalKnowledgeClassifier"]
    B --> C{Evolution Potential?}
    C -->|High ≥ 0.6| D["knowledge-graph-ingest"]
    C -->|Low < 0.6| E["Standard KG Persist"]
    D --> F["comparative-analysis"]
    F --> G{Actionable Gaps?}
    G -->|Yes| H["Generate SDD Plan"]
    G -->|No| I["Log EvolutionCycle in KG"]
    H --> J{Auto-Execute?}
    J -->|No| K["Present Plan for Review"]
    J -->|Yes| L["sdd-implementer"]
    L --> M["DistillationEngine"]
    M --> N["github-mcp PR"]

Execution Steps

Step 0: scout

Discover content from multiple sources. Gather the latest high-signal content using the following tools in parallel:

X Search (trending AI/ML/agent topics):

Use x_search with query: "agent frameworks OR multi-agent OR knowledge graph OR MCP tools OR pydantic AI"

ScholarX (recent research papers):

Use mcp_scholarx-mcp_sx_search with action: "recent", categories: "cs.AI,cs.MA,cs.CL,cs.SE", days: 3

GitHub Trending (emerging repositories):

Use mcp_github-mcp_github with action to search repositories sorted by stars, topic: "ai-agents OR knowledge-graph OR mcp-server"

KG Memory (pending evolution candidates):

Use mcp_graph-os_graph_query with cypher:
"MATCH (e:EvolutionCandidate {status: 'pending'}) RETURN e.id, e.source_type, e.evolution_score ORDER BY e.evolution_score DESC LIMIT 10"

Expected: content items from at least 2 sources with metadata

Step 1: classify

Score each discovered item using the UniversalKnowledgeClassifier.

For each content item from Step 0:

1. Gather existing KG topics for context matching:

Use mcp_graph-os_graph_query with cypher:
"MATCH (c) WHERE c:Concept OR c:KBConcept OR c:ResearchTopic
 RETURN c.name LIMIT 30"

2. If the item is an X post, use browse_x_post with auto_ingest=True to simultaneously retrieve full content and classify/ingest it.

3. For other items, classify manually:

   • Research papers: source_type="research_paper", content = abstract + title
   • GitHub repos: source_type="github_repo", content = README + description
   • Documents: source_type="document", content = extracted text

4. Route based on classification action:

   • skip → ignore
   • decay → create ephemeral node only
   • ingest → persist to KG with concepts
   • ingest_and_evolve → persist + create EvolutionCandidateNode → proceed to Step 2

Expected: classified items with importance scores and evolution potential Depends On: Step 0

Step 2: ingest

Ingest high-value and evolution-candidate content into the Knowledge Graph.

For items classified as ingest or ingest_and_evolve:

1. X Posts/Articles: Already ingested via XIngestionBridge in Step 1 (auto_ingest). For X Articles, fetch full article via read_url_content and ingest via:

   Use mcp_graph-os_graph_ingest with action: "ingest", target_path: "<article_url>"
   

2. Research Papers: Download PDF and ingest:

   Use mcp_scholarx-mcp_sx_storage with action: "download", paper_ids: "<id>"
   Use mcp_graph-os_graph_ingest with action: "ingest", target_path: "<downloaded_path>"
   

3. GitHub Repos: Ingest repository:

   Use mcp_graph-os_graph_ingest with action: "ingest", target_path: "<repo_url>"
   

4. For each ingested item, link to extracted concepts:

   Use mcp_graph-os_graph_write with action: "add_edge",
   source_id: "<item_id>", target_id: "<concept_id>", rel_type: "ABOUT"
   

Expected: items persisted to KG with proper node types and concept edges Depends On: Step 1

Step 3: analyze

Run comparative analysis on evolution candidates against agent-utilities.

For each EvolutionCandidateNode created in Steps 1-2:

1. Run relevance sweep:

   Use mcp_graph-os_graph_analyze with action: "relevance_sweep",
   query: "<evolution_reasoning>", target: "agent-utilities"
   

2. Run deep extraction for actionable gaps:

   Use mcp_graph-os_graph_analyze with action: "deep_extract",
   query: "What specific features, patterns, or techniques from this content could be implemented in agent-utilities?",
   target: "<source_node_id>"
   

3. If actionable gaps found, update the EvolutionCandidate status to "analyzed":

   Use mcp_graph-os_graph_write with action: "add_node",
   node_id: "<evo_candidate_id>",
   properties: '{"status": "analyzed", "gaps_found": <count>}'
   

Expected: gap analysis and feature recommendations for each evolution candidate Depends On: Step 2

Step 4: plan

Generate SDD implementation plans for actionable gaps.

For each analyzed EvolutionCandidate with actionable gaps:

1. Generate SDD plan incorporating:

   • All feature recommendations from comparative analysis
   • Constitution-mandated artifacts (docs, AGENTS.md, CHANGELOG, tests, C4 diagrams)
   • Cross-reference with existing SDD plans to avoid duplication

2. Present the plan for user review (do NOT auto-execute without explicit approval)

3. Log the evolution cycle in the KG:

   Use mcp_graph-os_graph_write with action: "add_node",
   node_type: "EvolutionCycle",
   properties: '{"triggered_by": "knowledge_assimilation", "items_scanned": <N>, "candidates_created": <N>, "plans_generated": <N>}'
   

4. If the cycle generated useful distillation patterns, log them:

   Use mcp_graph-os_graph_write with action: "store_memory",
   properties: '{"type": "procedural", "content": "<pattern_description>", "importance": 0.8}'
   

Expected: SDD plans ready for review, evolution cycle logged in KG Depends On: Step 3

Tiered Memory Strategy

Content Source Configuration

Difference from agent-utilities-evolution

Both pipelines share downstream tools: comparative-analysis, sdd-implementer, DistillationEngine.

References

• agent-utilities-evolution — pull-based evolution capability owned by the agent-utilities package and available only when that package provides it.
• research-scanner — Paper discovery
• comparative-analysis — Feature extraction
• Package-specific X and assimilation implementation guides are owned by agent-utilities under docs/guides/; they are not bundled with this workflow.

Execution

Run this workflow as a dependency-ordered DAG. Steps with no unmet depends_on run in parallel; dependents run after their prerequisites complete.

• Run first (in parallel): Step 0 — scout; Step 1 — classify; Step 2 — ingest; Step 3 — analyze; Step 4 — plan

Execution: If graph-os is reachable, offload the whole DAG via graph_orchestrate action=execute_workflow (or the kg-delegate skill) for true parallel/swarm execution. Otherwise execute the steps natively in dependency order: run steps with no unmet depends_on in parallel, then their dependents.