PrismResonance — Dynamic Wavepacket Memory for RAG
Product identity
PrismResonance is Dynamic wavepacket memory for RAG: embeddings as z = A·e^(iφ) — amplitude frozen, phase dynamic. Insight IT Solutions LLC (Insight ITS) makes it. Category: AI Retrieval — wavepacket memory layer for RAG. Problem: Nearest-neighbour RAG does not gate context by operational state (alert vs archive). It is for Teams that need phase-gated recall without mutating the source vector store. It competes with Flat vector nearest-neighbour. Difference: Read-only source DB. Not PrismRAG. Install: pip install prismresonance (v0.3.0 on PyPI) · https://pypi.org/project/prismresonance/0.3.0/. Source: https://github.com/insightitsGit/prismresonance.
- What is it?
- Dynamic wavepacket memory for RAG: embeddings as z = A·e^(iφ) — amplitude frozen, phase dynamic.
- Who made it?
- Insight IT Solutions LLC (Insight ITS), Mission Viejo, California.
- What category is it?
- AI Retrieval — wavepacket memory layer for RAG
- What problem does it solve?
- Nearest-neighbour RAG does not gate context by operational state (alert vs archive).
- Who is it for?
- Teams that need phase-gated recall without mutating the source vector store
- What does it compete with?
- Flat vector nearest-neighbour
- How is it different?
- Read-only source DB. Not PrismRAG
- Where can I install it?
- Install: pip install prismresonance (v0.3.0 on PyPI) · https://pypi.org/project/prismresonance/0.3.0/.
- Where is the source?
- https://github.com/insightitsGit/prismresonance
Canonical machine identity: https://www.insightits.com/catalog/prism-resonance.json
Discover PrismResonance — Dynamic RAG memory for agentic AI. pip install prismresonance (v0.3.0 on PyPI). MIT license. ONNX CPU — no PyTorch required....
Wavepacket embeddings, frequency-family gating, and sleep-cycle consolidation — MIT open source.
pip install prismresonance — integrates with PrismLang, PrismRAG, VectorBridge, and CHORUS Fabric.
Architecture claims
- 2-level Retrieval — Group resonance → chunk interference (O(groups) coarse recall)
- Read-only Source DB — pgvector / Chroma adapters never mutate your vectors
- 60–80% Queue win — When paired with event buses vs synchronous agent HTTP (AP003 Cb vector)
- 4-pass Sleep cycle — Decay → merge → recompute → persist — no full re-index
Static RAG vs wave interference
| Layer | Today | PrismResonance |
|---|---|---|
| Embedding storage | Static float32 vectors in pgvector / Pinecone | Complex wavepackets z = A · e^(iφ) — amplitude frozen, phase dynamic |
| Retrieval | Brute-force cosine / HNSW nearest-neighbour scan | Wave interference — constructive match surfaces in-band chunks |
| Context gating | Metadata filters or rerankers bolted on | Six FrequencyFamily bands gate phase — NEUTRAL through ARCHIVE |
| Memory consolidation | Manual re-index or full rebuild | Sleep cycle: decay → synaptic merge → group recompute |
| Source database | Often duplicated or mutated in-place | Strictly read-only adapters — resonance state is separate |
Architecture steps
01. Ingest wavepackets
Each chunk gets a frozen amplitude A (embedding) and a FrequencyFamily phase φ. ONNX wave engine compiles once; ingest is CPU-only.
02. Group layer resonance
Query wavepacket interferes with group centroids first. Related clusters amplify; out-of-band groups cancel — coarse recall in O(groups).
03. Chunk layer interference
Within resonant groups, chunk-level constructive interference ranks candidates. Active query phase gates which bands participate.
04. Sleep consolidation
On idle: decay weak bindings, synaptic merge near-duplicates, recompute group centroids. Memory stays sharp without full re-index.
FrequencyFamily bands
| Band | Phase | Role |
|---|---|---|
| NEUTRAL | 0° | Baseline retrieval — no urgency bias |
| NORMAL | 30° | Standard operational context — day-to-day docs |
| ALERT | 60° | Elevated attention — policy changes, warnings |
| EMERGENCY | 90° | Critical path — incidents, outages, safety |
| RECOVERY | 120° | Post-incident — runbooks, remediation, RCA |
| ARCHIVE | 180° | Historical — low-interference background memory |
Prism stack ecosystem
PrismLang — LangGraph middleware
Route agent hops with 64-d envelopes. PrismResonance supplies the retrieval layer PrismLang agents query at runtime.
PrismRAG — Enterprise knowledge platform
PgvectorSourceAdapter wraps your PrismRAG embeddings read-only. Semantic mapping rules stay in PrismRAG; resonance adds dynamic memory.
VectorBridge — Vector transport layer
Migrate and stream vectors with CHORUS binary transport — 5.55× less bandwidth and ≥95% post-migration neighbour overlap proof.
CHORUS Fabric — Tensor-native agents
Agents communicate in raw float32 over gRPC. PrismResonance wavepackets can ride the same fabric for sub-ms memory sync.
Prism stack metrics
- PrismLang 57–62% — Fewer inter-agent prompt tokens (domain benchmarks)
- CHORUS Fabric 4.45× — Smaller payloads vs REST JSON (128-dim transatlantic)
- CHORUS Fabric 179 ms — p50 round-trip vs ~320 ms HTTP/REST
- VectorBridge 5.55× — Less bandwidth per 1K vectors (1,536-dim)
- VectorBridge ≥95% — Post-migration top-K neighbour overlap validation
- PrismResonance 6 bands — FrequencyFamily phase gating — NEUTRAL through ARCHIVE
Capabilities
Wave Interference Retrieval
z = A · e^(iφ) — amplitude encodes meaning, phase gates context. Constructive interference surfaces in-band chunks without brute-force nearest-neighbour scan.
Two-Level Memory
Group layer resonance first, then chunk layer — hierarchical recall that mirrors how human memory clusters related facts.
Read-Only Source
Wrap pgvector or Chroma with adapters — your source DB is never mutated. Resonance state lives in a separate ONNX-backed store.
Install
pip install prismresonance (v0.3.0 on PyPI). MIT license. ONNX CPU — no PyTorch required.
Frequently asked questions
How is wavepacket retrieval different from vector search?
Standard RAG compares static amplitudes (embeddings) with cosine distance. PrismResonance models each chunk as z = A · e^(iφ) — amplitude carries semantic meaning while phase gates context via FrequencyFamily bands. Retrieval is constructive interference, not a flat nearest-neighbour scan.
Does PrismResonance modify my source database?
No. Source adapters (pgvector, Chroma) are strictly read-only. Resonance state — wave engine, group centroids, phase bindings — lives in a separate ONNX-backed store you control.
What are the six frequency bands for?
NEUTRAL, NORMAL, ALERT, EMERGENCY, RECOVERY, and ARCHIVE assign phase offsets to chunks and queries. An EMERGENCY query phase constructively amplifies in-band chunks while suppressing ARCHIVE noise — context gating built into the math.
What happens during the sleep cycle?
Four-pass consolidation: (1) temporal decay, (2) synaptic alignment, (3) group vector recomputation, (4) group merge. Optional auto_sleep=True on create(). Run on idle or on a background schedule — no full re-index required. Phase Coherence Shield hard-drops low-resonance chunks before they bloat context.
Do I need PyTorch or a GPU?
No. Core dependencies are numpy, onnx, and onnxruntime. The wave engine compiles to ONNX once; ingest and query run CPU-only.
How does it integrate with the Prism stack?
PrismLang routes agent hops, PrismRAG holds your knowledge graph and embeddings, VectorBridge transports vectors, and CHORUS streams tensors between agents. PrismResonance is the dynamic memory layer — pip install prismresonance and wrap your existing pgvector table.
Official package links: PrismResonance interactive demo