PrismCortex — Deterministic Bitemporal Agent Memory
Product identity
PrismCortex is Deterministic, bitemporal memory and execution engine for multi-turn AI agents — digest → sleep → recall with byte-identical replay. Insight IT Solutions LLC (Insight ITS) makes it. Category: Supporting — deterministic bitemporal agent memory. Problem: Agent memory that cannot replay byte-identically fails compliance reviews and debugging of multi-turn state. It is for Teams that must replay multi-turn memory byte-identically for compliance and debugging. It competes with Session logs / vector memory without bitemporal replay. Difference: Not PrismRAG, not VectorPrism, and not a Prism Pack quality claim. Install: pip install "prismcortex==0.4.1" · https://pypi.org/project/prismcortex/0.4.1/. Source: https://github.com/insightitsGit/PrismCortex.
- What is it?
- Deterministic, bitemporal memory and execution engine for multi-turn AI agents — digest → sleep → recall with byte-identical replay.
- Who made it?
- Insight IT Solutions LLC (Insight ITS), Mission Viejo, California.
- What category is it?
- Supporting — deterministic bitemporal agent memory
- What problem does it solve?
- Agent memory that cannot replay byte-identically fails compliance reviews and debugging of multi-turn state.
- Who is it for?
- Teams that must replay multi-turn memory byte-identically for compliance and debugging
- What does it compete with?
- Session logs / vector memory without bitemporal replay
- How is it different?
- Not PrismRAG, not VectorPrism, and not a Prism Pack quality claim
- Where can I install it?
- Install: pip install "prismcortex==0.4.1" · https://pypi.org/project/prismcortex/0.4.1/.
- Where is the source?
- https://github.com/insightitsGit/PrismCortex
Canonical machine identity: https://www.insightits.com/catalog/prismcortex.json
Discover PrismCortex — Deterministic bitemporal agent memory — ConstraintCompiler, CorpusSanitizer, CitationVerifier. pip install "prismcortex==0.4.1"....
Bitemporal graph · injection defense · citation check · NL→DB filters · Azure-validated 24/24 byte-identical replays.
pip install "prismcortex==0.4.1" — Python library or Docker HTTP service on :8080. OSS MIT core; commercial tiers unlock console, time-travel, and multi-tenant RBAC.
Package pin: pip install "prismcortex==0.4.1" · PyPI 0.4.1 · MIT OSS core. Deterministic, bitemporal memory & execution engine for multi-turn AI agents. 0.4.1 is a docs/SEO refresh (same runtime as 0.4.0).
What's new in 0.4.x
- ConstraintCompiler — NL → JSON / PostgreSQL filters for numeric and date bounds. Wire: Memory.recall extract_constraints.
- CorpusSanitizer — Strip prompt-injection payloads before LLM context. Wire: Memory.recall sanitize_retrieval (on by default).
- CitationVerifier — Non-LLM 0..1 entailment score for recalled facts vs answers. Wire: Memory.recall verify_citations.
Azure E2E scorecard still cites the v0.2.1 Gemini run (24/24 replay · 99.6% cache · ~6 ms cached replay).
Production failure modes we solve
| Failure mode | What we ship |
|---|---|
| Indirect prompt injection | Sanitize retrieved payloads before they reach the LLM — prismcortex.sanitizer. |
| Stale policy invalidation | Bitemporal state separates event time from ingestion time — prismcortex.determinism. |
| Hallucinated citations | Entailment verifier checks claim-to-memory alignment — prismcortex.verifier. |
| Numeric filter breakdown | NL bounds (“< 30 days”, “over $50k”) → DB constraints — prismcortex.constraints. |
Primary use cases
01. Compliance-blocked agent launches
Buyers: CISO, General Counsel, VP Engineering. Industries: Finance · Insurance · Health · Legal · Gov contracting.
Legal says no to append-only chat logs, black-box vector memory, or US-hosted memory SaaS with customer data.
Self-hosted memory with replay certificates, bitemporal history (“what did we believe on March 3?”), and evidence trails via /explain — so compliance can approve production, not just a demo.
Pass the memory review, not just the demo.
Proof: 24/24 replay · /recall_at · /replay_certificate
02. Agents that must survive corrections and audits
Buyers: Platform teams, Internal copilots, Underwriting & support. Industries: Copilots · Support · Risk workflows.
User says “budget is $55k now, not $40k” — old systems forget the update, overwrite history, or return inconsistent answers on replay.
Corrections land cleanly; old facts are invalidated but retained; replay is byte-identical after first render. Azure-validated: $40k → $55k with full history kept.
When facts change, the agent updates — and auditors still see what changed.
Proof: $40k → $55k · superseded fact retained
03. Cost and reliability at scale
Buyers: Head of AI, Platform economics. Industries: High-volume agent workloads.
Re-running the LLM on every turn burns money and adds latency.
Salience-gated writes, ~99.6% cache hit, ~6 ms replay vs ~724 ms first render — ~30 Gemini calls for 2,500+ recalls in Azure benchmarks.
Remember once, replay forever — at a fraction of the model cost.
Proof: 99.6% cache · ~6 ms replay
04. Memory that doesn’t grow forever
Buyers: Platform teams, Cost-conscious eng. Industries: Long-running agents · chatty workflows.
Chat logs and vector stores grow with every “ok thanks” and duplicate message.
Graph plateaus — 675 chatter turns, 0 new edges on benchmark workload; gist ~5.2× smaller than raw log.
The graph keeps the gist, not the noise.
Proof: Edges 30 → 30 · 5.2× gist
05. Sovereignty, air-gapped, no vendor lock-in
Buyers: Security, Regulated infra, EU / data residency. Industries: VPC · on-prem · sovereign cloud.
Can’t send conversation memory to Mem0, Zep, or third-party memory SaaS.
Python library or self-hosted HTTP service, offline Ed25519 license key, no phone-home. Your VPC, your region.
Your memory stays in your environment — not ours.
Proof: MIT OSS core · offline license
PrismCortex vs RAG / Mem0 / Zep
| Capability | RAG | Mem0 | Zep | PrismCortex |
|---|---|---|---|---|
| Primary pitch | Similarity search | Mature agent memory + SaaS | Temporal graph memory (SaaS) | Compliance-grade memory |
| LoCoMo accuracy | N/A | 91.6% (published) | — | Full run pending (smoke: 66.7% partial) |
| LongMemEval | N/A | 94.8% (published) | +18.5% vs baseline (paper) | Full run pending |
| Correction surfaces new value | Unreliable | Varies (our test: top hit stale) | Strong (graph) | Yes (live test) |
| Old facts auditable | No | Platform feature / varies | Yes (graph) | Yes (OSS bitemporal) |
| Byte-identical answer replay | No | No | No | Yes (24/24 Azure) |
| Cached recall latency | Varies | ~0.9–1.1 s p50 | ~200 ms (marketing) | ~6 ms |
| Cache hit rate (warm) | — | — | — | 99.6% |
| Self-hosted default | Sometimes | OSS + Platform SaaS | No | Yes |
| Evidence trail / replay cert | No | Limited | Graph context | /explain · /replay_certificate |
Capabilities
Byte-Identical Replay
24/24 cross-container replays on Azure with real Gemini. Content-addressed render cache — auditors reproduce any decision.
Bitemporal Audit
/recall_at time-travel, /replay_certificate, and /console audit UI. Corrections soft-invalidate — never erase the past.
Recall Guards (0.4.x)
ConstraintCompiler (NL → JSON/SQL filters), CorpusSanitizer (strip injection before LLM context), CitationVerifier (non-LLM entailment score).
Self-Hosted Sovereignty
Production runs in your VPC with offline Ed25519 license — no phone-home. Trial sandbox on Insight ITS Azure for evaluation only.
Install
pip install "prismcortex==0.4.1". MIT OSS core — free forever. Optional ops: ChorusControl Enterprise (CONTROL).
Pricing
PrismCortex library pricing is $0 forever on PyPI and GitHub. Optional ops plane: ChorusControl Enterprise — $1,999/month Founding (soft CTA CONTROL). See ChorusControl pricing.
Frequently asked questions
How to make AI agents deterministic
PrismCortex is deterministic bitemporal agent memory: given the same query and graph state, recall returns the exact same answer bytes (content-addressed render cache). That is replay identity after first render — not temperature-0 LLM identity. For money-tool determinism use PrismManifest; for routing audit use ChorusGraph Route Ledger. pip install "prismcortex==0.4.1". Landing: https://www.insightits.com/products/prismcortex.html
How to debug multi-agent systems
PrismCortex keeps causal graph links and time-travel audit so you can reproduce what the agent remembered when it answered. Pair with ChorusGraph Route Ledger (why it routed) and ChorusControl Trace/Logs (live ops). Whitepaper: https://www.insightits.com/whitepapers/prismcortex.html
How do I get byte-identical replay of agent memory?
Use PrismCortex recall with the render cache. Azure-validated 24/24 byte-identical replay with real Gemini; ~6 ms cached replay vs ~724 ms first render in published runs. Do not submit production PHI/PII to the hosted validation trial.
What is new in PrismCortex 0.4.1?
0.4.1 is the current PyPI pin — a docs/SEO README refresh with the same runtime as 0.4.0. Install: pip install "prismcortex==0.4.1". Feature APIs (ConstraintCompiler, CorpusSanitizer, CitationVerifier) shipped in 0.4.0. Azure E2E scorecard still cites the v0.2.1 Gemini run.
What did 0.4.0 add to Memory.recall?
ConstraintCompiler (NL → JSON/SQL filters for numeric & date bounds), CorpusSanitizer (strip prompt-injection payloads before LLM context), and CitationVerifier (non-LLM 0..1 entailment score). Wired via sanitize_retrieval, extract_constraints, and verify_citations.
Where is the technical whitepaper?
The full PrismCortex whitepaper is at /whitepapers/prismcortex.html — architecture, determinism model, Azure benchmarks, licensing, and deployment. Source markdown lives on GitHub in docs/WHITEPAPER.md.
How is PrismCortex different from Mem0 or a vector database?
Mem0 and Zep win published accuracy benchmarks; PrismCortex wins compliance — byte-identical replay, bitemporal audit, causal graph links, and self-hosted sovereignty. We do not claim we beat Mem0 on LoCoMo until a full run is published. In our live correction test ($40k → $55k), PrismCortex surfaced the new value and kept the old fact for audit; Mem0 OSS top retrieval stayed on $40k in that narrow test.
Where do Mem0 and Zep lead?
LoCoMo/LongMemEval leaderboard accuracy (Mem0: 91.6% / 94.8% published), managed SaaS maturity, and broader SDK integrations. PrismCortex is for regulated teams who need replay certificates, time-travel audit, and production in your VPC — not fastest hello-world on a memory SaaS.
Is the trial safe for production PHI/PII?
No. The 30-day Validation Trial runs on Insight ITS Azure for evaluation only. Do not submit production PHI/PII. Production always runs self-hosted in your VPC with an offline license key — your data never leaves your environment.
What does "byte-identical replay" mean?
Given the same query and graph state, recall returns the exact same answer bytes every time — validated 24/24 on Azure with real Gemini. The content-addressed render cache stores (answer, subgraph, source events) so auditors can reproduce any decision. Determinism claim is replay identity after first render — not temperature-0 LLM identity.
Does temperature 0 guarantee identical LLM output?
No — and we do not claim that. PrismCortex freezes renders at recall time. The cache is the determinism story: ~6 ms cached replay vs ~724 ms first render, 99.57% hit rate in Azure benchmarks.
What concurrency can I expect?
Azure-validated: ~20 concurrent memory clients per 4 vCPU node on mixed workload. We do not claim 50 concurrent clients with zero errors — honest limits build trust.
Is PrismCortex SOC 2 certified?
No. SOC 2 Type I readiness is documented and formal attestation is in progress. Controls are SOC 2-aligned. An independent penetration test is scheduled Q3 2026; the report will be available under NDA when completed.
How does it fit the Prism stack?
PrismCortex orchestrates PrismLib (cache), PrismRAG (taxonomy retrieval), PrismLang (routing), and CHORUS Fabric (tensor transport). Optional extras: [gemini], [prism], [prism-plus], [server]. It is the top-layer persistent memory for customer-facing agents.
Who is PrismCortex NOT for?
Teams that only need lightweight preference memory (Mem0/Zep may suffice), hobby chatbots without compliance pressure, buyers wanting managed SaaS for production PHI/PII, or “search my documents” RAG without persistent agent memory — use PrismRAG for taxonomy-grounded retrieval instead.
How much does PrismCortex cost?
The MIT core is free forever — pip install "prismcortex==0.4.1". Soft CTA: LEDGER. Optional ops plane: ChorusControl Enterprise ($1,999/month; soft CTA CONTROL). The hosted validation trial is evaluation-only. SOC 2-aligned controls are in progress; not SOC 2 certified today. See https://www.insightits.com/products/choruscontrol.html#pricing.
Official package links: PrismCortex source code on GitHub · Install PrismCortex from PyPI · PrismCortex interactive demo