PrismCortex

Deterministic, bitemporal memory engine for multi-turn AI agents.

Eliminate context decay, stale vector collisions, non-deterministic replay failures, and hallucination ratchets in production RAG/agent architectures.

Repository: https://github.com/insightitsGit/PrismCortex (public) · Package: prismcortex 0.4.0

Author: Amin Parva · Company: Insight IT Solutions LLC · www.insightits.com

AI agent handoff · Whitepaper · Use cases · Benchmarks · How we compare · Design

Product page: insightits.com/products/prismcortex

---

Why standard RAG memory fails

Failure mode	What goes wrong in production
Stale knowledge collisions	Naive vector similarity returns superseded policies alongside active ones — the agent cites both.
Context decay & anaphora loss	Parent entity subject is lost across multi-turn sessions; “it / that / the policy” no longer resolve.
Multi-hop blindness	Continuous vector spaces fail at logical dependency joins (A→B→C) that a graph can walk.
Non-reproducible replays	You cannot prove byte-identical state for audit — temperature-0 LLM calls still drift.

---

How PrismCortex fixes it

Capability	Where it lives	What you get
Bitemporal auditing	determinism.py, graph edges (valid_from / valid_to)	Separates real-world event time from ingestion/system time; corrections soft-invalidate, never erase.
Causal graph links	engine.py, tests/test_graph_engine.py	Extracted facts become relational edges — not isolated float arrays.
Salience & consolidation	salience.py, Memory.sleep()	Skips low-value turns; parks uncertain facts; consolidates without dropping history.
Byte-identical replay	content-addressed cache + /replay_certificate	Reproducible answer audits for SOC 2–aligned / compliance workflows.

API surface (real): digest() → graph · recall() → frozen answer · sleep() → consolidate · explain() → evidence.

---

5-line quickstart

from prismcortex import reference_memory

mem = reference_memory()  # GEMINI_API_KEY for real extraction
mem.digest("California parental leave updated to 12 weeks.")
print(mem.recall("What is our CA leave policy?").answer)
# Correction keeps history: digest a change → recall new value; old edge retains valid_to.

Zero-dependency demo (no API key — rule-based extractor):

python examples/quickstart.py

With real Gemini:

pip install "prismcortex[gemini]"
GEMINI_API_KEY=... python examples/quickstart.py

---

Competitive positioning

Mem0 / Zep lead published accuracy suites (LoCoMo, LongMemEval). PrismCortex leads compliance — temporal audit, consolidation, causal graph, and byte-identical replay. Do not claim LoCoMo wins until a full PrismCortex run is published.

Capability	Naive vector RAG	Mem0 / Zep	PrismCortex
Temporal auditing	No (append / re-rank)	Platform / graph varies	Yes — OSS bitemporal edges
Execution replay	No	No byte-identical answers	Yes — 24/24 Azure E2E
Memory consolidation	Truncate / summarize	Product features vary	Yes — salience + sleep()
Causal graph engine	Flat embeddings	Zep: temporal graph; Mem0: memory store	Yes — relation edges + evidence

Live correction head-to-head (same Gemini): PrismCortex surfaced $55k after $40k → $55k; Mem0 OSS top retrieval stayed $40k in our run — vs_mem0.json.

Full tables: compare.md · docs/COMPETITIVE.md

---

Validated claims (Azure E2E, real Gemini, v0.2.1)

Claim	Result
Replay determinism	24/24 byte-identical
Corrections + audit	$40k → $55k; superseded fact retained
Cache	99.6% hit — 30 Gemini / 2,563 recalls
Cached replay	~6 ms vs ~724 ms first render
Mixed load (c=20)	0 errors on 4 vCPU · slo_pass: true
Scale (50k facts, ANN)	85% hit@8, 74 ms p95

Details: benchmarks/RESULTS.md

---

Install

pip install prismcortex                  # core (MIT)
pip install "prismcortex[gemini]"        # + real Gemini extraction/rendering
pip install "prismcortex[prism]"         # + Insight ITS stack with prismlib
pip install "prismcortex[prism-plus]"    # + same stack with prismlib-plus
pip install "prismcortex[server]"        # + FastAPI HTTP service

Requires Python 3.10+. [prism] and [prism-plus] are mutually exclusive.

HTTP service

export GEMINI_API_KEY=...
export PRISMCORTEX_API_KEY=your-secret
uvicorn prismcortex.server:app --host 0.0.0.0 --port 8080
# OpenAPI: http://localhost:8080/docs

What's new in 0.4.0

• ConstraintCompiler — NL → JSON / PostgreSQL filters for numeric & date bounds
• CorpusSanitizer — strip prompt-injection payloads before LLM context
• CitationVerifier — non-LLM 0..1 entailment score for recalled facts vs answers
• Wired into Memory.recall (sanitize_retrieval, extract_constraints, verify_citations)
• Release notes: docs/CHANGELOG_0.4.0.md

What's new in 0.3.0

• mem.on_event(callback) — correction / conflict / forget (MemoryEvent)
• Evidence fields: valid_from, supersedes_prior, prior_value
• [prism-plus] extra — see docs/CHANGELOG_0.3.0.md

---

Architecture

digest(text) ─▶ salience gate ─▶ extract gist ─▶ delta in RAM
                   ├─ certain / urgent ─▶ commit  (version++)
                   └─ uncertain ───────▶ staging ──▶ sleep() ──▶ commit

recall(query) ─▶ retrieve subgraph ─▶ cache hit? replay (byte-identical)
                                    └─ miss? render once → freeze

Determinism claim (honest): we do not claim temperature-0 LLM identity. We claim replay identity after first render for a (query, memory-version) pair. See DESIGN.md.

---

Development

git clone https://github.com/insightitsGit/PrismCortex.git
cd PrismCortex
pip install -e ".[dev,gemini,server]"

pytest tests/                          # graph tests need no API key
GEMINI_API_KEY=... pytest              # full suite
python examples/quickstart.py          # zero-deps path
python benchmarks/scale_bench.py --ann

Azure E2E / load driver (needs running server): python benchmarks/driver.py — see docs/OPS_RUNBOOK.md.

---

Documentation

Doc	Contents
AGENTS.md	Canonical URLs, contacts, processes
docs/USE_CASES.md	Problem → architecture mappings
docs/WHITEPAPER.md	Product whitepaper
DESIGN.md	Engineering design
benchmarks/RESULTS.md	Azure scorecard
compare.md / docs/COMPETITIVE.md	Market comparison
docs/CAPACITY.md / docs/LOAD_BENCHMARK.md	Sizing & load SLO
SECURITY.md	Security posture

---

Licensing

Open-core (MIT): digest / recall, bitemporal graph, determinism cache — free on PyPI.

Commercial: audit console, advanced governance, scale tiers — offline Ed25519 key, no phone-home.

Enterprise: info@insightits.com · +1 (973) 692-6919 · Insight IT Solutions LLC · www.insightits.com

Author: Amin Parva