DEV Community

Max Klein
Max Klein

Posted on

How spreading activation beats keyword search for LLM agents

#rust #llm #ai #opensource

LLM agents are powerful reasoners but poor navigators.

They can analyze code, write solutions, and debug problems. But ask them "what does this change affect?" or "what am I missing?" and they're stuck grepping through flat files.

I built m1nd to fix this.

The idea

Instead of searching your data, m1nd activates it. You query a concept, and the graph lights up: connected nodes fire with decaying signal, noise gets cancelled, and the system learns from your feedback.

You ask:  "What's related to authentication?"
m1nd:     Activates auth -> propagates to session, middleware, JWT
          -> detects structural hole (no rate limiter connected)
          -> predicts co-change with password reset module
          -> all in one query, ranked by multi-dimensional relevance
Enter fullscreen mode Exit fullscreen mode

How it works

1. Ingest - Build a property graph from your data (code, JSON, any domain)

2. Activate - Spreading activation across 4 dimensions:

  • Structural: graph topology, PageRank, community structure
  • Semantic: label similarity via char n-grams and co-occurrence
  • Temporal: recency decay, change velocity, co-change history
  • Causal: directed dependency flow along import/call/contain edges

3. Cancel noise - XLR differential processing maintains parallel signal and noise graphs, gating activation through an adaptive sigmoid. Borrowed from balanced audio engineering.

4. Learn - Hebbian plasticity: tell m1nd "this result was correct" and it strengthens the connections (LTP). "This was wrong" weakens them (LTD). Over time, the graph gets smarter.

5. Persist - Graph + plasticity state saves to disk. Next session starts where this one left off.

The 13 tools

m1nd exposes 13 tools over MCP (Model Context Protocol), callable by any LLM agent:

Category Tools
Discovery activate, why, missing, fingerprint, resonate
Change Analysis impact, predict, counterfactual
Learning learn, drift, warmup
System ingest, health

Real example

m1nd analyzing its own codebase (693 nodes, 2007 edges):

> m1nd.activate query="spreading activation"
0.94 fn::propagate_wavefront  (structural 0.97)
0.91 fn::score_candidates     (semantic 0.93)
0.87 fn::xlr_gate             (causal 0.89)
XLR cancelled 12 noise edges

> m1nd.counterfactual remove=["graph.rs"]
CASCADE: 31 nodes lose connectivity
ORPHANED: 8 functions become unreachable
graph.rs is load-bearing
Enter fullscreen mode Exit fullscreen mode

Why not vector search?

Vector search finds things that look similar. Spreading activation finds things that are structurally connected. The difference matters:

  • Blast radius requires graph topology, not cosine similarity
  • Structural holes are invisible to embeddings
  • Co-change prediction needs temporal graph history
  • Counterfactual simulation is impossible without structure
  • Hebbian learning is online and incremental - no retraining

Try it

Built in Rust. 15.5K LOC. 159 tests. 3.8MB binary.

cargo build --release
./target/release/m1nd-mcp
Enter fullscreen mode Exit fullscreen mode

GitHub: github.com/maxkle1nz/m1nd
Landing: maxkle1nz.github.io/m1nd
Integration guide: INTEGRATION-GUIDE.md

*Max Kleinschmidt *

Top comments (0)