Agent memory that
improves itself

The dual loop in action: PRISM composes scenarios, runs them against Graphonomous, judges the results across 9 CL dimensions, then reflects on what to test next. Meanwhile, Graphonomous stores every evaluation result as knowledge nodes — so each cycle starts with richer context than the last.

Cycle 5 was the breakthrough: adversarial nodes with low confidence (0.29) outranked correct nodes (0.80+) because BM25 keyword scoring wasn’t confidence-weighted. PRISM caught it. We fixed it. Cycle 5c confirmed the fix. That’s the loop working.

Two loops, one inside the other:

PRISM (outer):  compose → interact → observe → reflect → diagnose
                              ↓
            ┌———————————— Graphonomous (inner) ————————————┐
            │ retrieve → route → act → learn → consolidate │
            └——————————————————————————————————————————————┘

You can do this with your own repo. PRISM’s BYOR (Bring Your Own Repo) registers your git history as ground truth, auto-discovers CL events from commits, and generates evaluation scenarios. Run it in Claude Code, Codex, or any MCP-capable agent.

Dual-loop interlocking with PRISM: When PRISM (OS-009) benchmarks Graphonomous, both closed loops nest. PRISM’s 6 machines + Graphonomous’s 5 machines = 11 tools in a shared session, down from 76. The outer loop improves the benchmark. The inner loop improves the memory. Each makes the other sharper.

MCP config: Add this to your .mcp.json or IDE settings to connect both servers:

{
  "mcpServers": {
    "graphonomous": {
      "command": "path/to/graphonomous/scripts/graphonomous_mcp_wrapper.sh",
      "args": ["--db", "~/.graphonomous/knowledge.db"]
    },
    "prism": {
      "command": "path/to/PRISM/scripts/prism_mcp_wrapper.sh",
      "args": []
    }
  }
}

Any system can add confidence scores. Any system can add a consolidation step. The difference is what happens when these features talk to each other.

In Graphonomous, a failed outcome reduces confidence on the causal parent nodes that informed the decision. Lower confidence changes the κ topology of that subgraph. Changed topology changes the routing decision on the next retrieval. The attention engine re-prioritizes goals based on the new coverage landscape. During idle time, consolidation prunes the weakened nodes and strengthens what worked. One outcome ripples through the entire system — retrieval, routing, attention, and memory lifecycle — without any component knowing about the others.

That causal chain breaks if any piece lives in a separate system. Bolted-on confidence scores don’t feed into topology analysis. A standalone consolidator can’t see which nodes were used in failed decisions. Separate goal tracking can’t query the graph’s epistemic coverage. The integrated architecture is the product.

• Letta (MemGPT) pioneered “continual learning in token space” with sleep-time consolidation and skill caching. But memory items have no confidence scores, no outcome-based feedback, and no topology awareness. Learning means rewriting text, not updating a causal graph.
• Zep / Graphiti builds temporal knowledge graphs with sophisticated entity deduplication and a four-timestamp bi-timeline. Strong on “what changed when.” But no confidence tracking, no outcome learning, no topology analysis, and contradictions are resolved by temporal invalidation alone — new overwrites old.
• Hindsight has an opinion network with updateable confidence — the closest competitor architecturally. But updates are linear alpha adjustments, not formal belief revision. No causal attribution chains, no topology routing, no consolidation cycles, and confidence exists only on opinions, not on facts or experiences.
• Cognee extracts knowledge graphs with a behavioral critic that scores memory by downstream utility — a genuine innovation. But no per-node confidence, no causal feedback loop, and no topology-aware retrieval.
• MemPalace achieves 96.6–100% retrieval accuracy with zero API calls — best-in-class for pure recall. But it’s storage and retrieval only. No learning loop, no confidence, no consolidation.

Graphonomous is the only system where confidence tracking, causal attribution, κ-routing, belief revision, sleep-cycle consolidation, multi-timescale memory, and goal-aware attention all operate on the same graph. MCP-native, works with any model, runs on SQLite at the edge.

GDPR-compliant forgetting is built in. Soft forget, cascade delete, policy-based pruning, and permanent audit-logged erasure (Article 17) — all in one tool surface.