PrismCortex

Deterministic, auditable, self-consolidating memory for AI agents.

Compliance-grade memory for regulated teams: byte-identical replay, bitemporal audit, and self-hosted sovereignty — not another vector chat log.

Repository: https://github.com/insightitsGit/PrismCortex (public)

🤖 AI agent handoff · 📄 Whitepaper · 📊 Benchmarks · ⚖️ How we compare · 🗺️ Roadmap · 🏗️ Design spec

Product page: insightits.com/products/prismcortex

---

AI assistants: docs/ai-overview.md · docs/llm-context.md · docs/architecture.md

What is this?

Deterministic, auditable, self-consolidating memory for AI agents (byte-identical replay, bitemporal audit).

Package: prismcortex 0.3.0 · Azure E2E scorecard still cites the v0.2.1 run

Who is it for?

Regulated teams needing compliance-grade agent memory, not a chat-log vector store.

What problem does it solve?

Chat-log / SaaS memory fails audit, correction, and residency requirements.

When NOT to use it

You only need ephemeral chat history with no audit requirements.

What's new in 0.3.0

• mem.on_event(callback) — correction / conflict / forget notifications for PrismShine and cache invalidation (MemoryEvent)
• Evidence correction metadata — valid_from, supersedes_prior, prior_value on /explain
• [prism-plus] extra — use prismlib-plus instead of prismlib (mutually exclusive with [prism])
• Release notes: docs/CHANGELOG_0.3.0.md

---

Why PrismCortex exists

Most agent memory is an append-only chat log or a vector store in someone else's cloud. That breaks in production when:

• Legal asks "what did the agent know on March 3rd?" — and you grep chat logs
• A correction ($40k → $55k) doesn't reliably surface — or erases audit history
• Compliance rejects third-party memory SaaS for data residency

PrismCortex digests each turn into a knowledge graph, consolidates uncertain facts in the background (sleep()), and recalls by rendering facts once and freezing answers in a content-addressed cache.

from prismcortex import reference_memory

mem = reference_memory(cache_path=".prismcortex_cache/demo.json")

mem.digest("My production deploy budget is $40,000.")
print(mem.recall("What's my deploy budget?").answer)        # → "$40,000"

mem.digest("Correction: my deploy budget is now $55,000.")  # fast-tracked (ALERT)
print(mem.recall("What's my deploy budget?").answer)        # → "$55,000"
# The $40,000 fact is still on record — time-stamped — for audit / time-travel.

---

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

Claim	Result
Replay determinism	24/24 byte-identical replays
Corrections + audit	$40k → $55k; superseded fact retained
Cost / cache	99.6% hit rate — 30 Gemini calls / 2,563 recalls
Cached replay	~6 ms vs ~724 ms first render
Mixed load (c=20)	0 errors on 4 vCPU node
Reference load SLO	PASS (slo_pass: true) — recall + mixed @ c=20, digest @ c=16
Server reliability	0 errors on core path
Scale (50k facts, ANN)	85% hit@8, 74 ms p95 retrieval

Details: benchmarks/RESULTS.md · docs/WHITEPAPER.md

---

How we compare

Mem0 and Zep lead published accuracy benchmarks (LoCoMo, LongMemEval, DMR). PrismCortex leads compliance — byte-identical replay, bitemporal audit, and self-hosted sovereignty.

	Mem0 (published)	Zep (published)	PrismCortex (live)
LoCoMo accuracy	91.6%	—	Full run pending
Correction test ($40k→$55k)	Top hit stale in our OSS run	—	Yes — new value + audit trail
Byte-identical replay	No	No	24/24 on Azure
Bitemporal audit (OSS)	Varies	Graph	Yes
Self-hosted default	OSS + SaaS	SaaS	Yes

Head-to-head: same Gemini, same correction — PrismCortex surfaced $55k after update; Mem0 OSS top retrieval stayed $40k in our live test. Reproducible: benchmarks/results/competitive/vs_mem0.json.

Landing page spec for agents: compare.md · Full technical comparison: docs/COMPETITIVE.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 (ChorusGraph)
pip install "prismcortex[server]"        # + FastAPI HTTP service
pip install "prismcortex[gemini,server,prism]"   # production stack

Requires Python 3.10+.

[prism] and [prism-plus] are mutually exclusive — both install the prism import namespace. Use [prism] for standalone PrismCortex; use [prism-plus] when the host already depends on prismlib-plus (e.g. ChorusGraph). Do not install both extras.

---

Two ways to run

1. Python library (in-process)

Best for a single agent embedded in your app:

from prismcortex import reference_memory

mem = reference_memory()   # needs GEMINI_API_KEY for real extraction
mem.digest("We use Postgres 16 in us-east-1.")
result = mem.recall("Where is our database hosted?")
print(result.answer, result.cache_hit, result.confidence)

# Optional: subscribe to corrections (PrismShine / semantic-cache eviction)
unsub = mem.on_event(lambda ev: print(ev.kind, ev.old_value, "→", ev.new_value))
# unsub() when done

2. HTTP service (multi-agent, Docker, Azure)

Best for platform teams and non-Python clients:

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

curl -X POST http://localhost:8080/digest \
  -H "Content-Type: application/json" \
  -H "X-API-Key: your-secret" \
  -d '{"text": "Our deploy budget is $40,000."}'

curl -X POST http://localhost:8080/recall \
  -H "Content-Type: application/json" \
  -H "X-API-Key: your-secret" \
  -d '{"query": "What is our deploy budget?"}'

Docker + Azure deploy: see deploy/run_only.sh.

---

Why it's different

	Append-only RAG	PrismCortex
Storage	every chat turn	graph topology (the gist)
Updates	append + hope retrieval ranks it	bitemporal: invalidate old, add new, keep history
Determinism	logs + LLM drift	content-addressed cache, replay-identical
Cost	re-extract every call	salience-gated writes, cached reads
Audit	grep the logs	evidence trail + replay certificate

---

Enterprise features (v0.3)

Feature	Endpoint / module
Explainability	POST /explain
Time-travel recall	POST /recall_at
Replay certificate	GET /replay_certificate
Conflict surfacing	GET /conflicts, POST /conflicts/resolve
GDPR erasure	POST /forget
Legal hold	POST /legal_hold
Multi-tenant + RBAC	auth.py, tenant.py
Audit console	GET /console
Metrics / ops	GET /metrics, GET /dashboard
50k+ facts (ANN)	PRISMCORTEX_USE_ANN=1
Correction events	Memory.on_event → MemoryEvent (library)

Docs: docs/SLA.md · docs/CAPACITY.md · docs/SOC2_ROADMAP.md · SECURITY.md

---

Architecture

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

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

Port	Reference (core, no Prism deps)	Production extras
Gist projection	hashing embeddings	prismlang ([prism] / [prism-plus])
Graph store	in-memory bitemporal	Cortex-owned store (+ prismrag-patch governor)
Consolidation	in-process	prismresonance
Render cache	JSON file	prismlib or prismlib-plus
Extraction	—	Gemini ([gemini])

Dependency note: pip install prismcortex needs only pydantic, numpy, and cryptography. Prism-family packages are optional via [prism] or [prism-plus].

Full design: DESIGN.md · Whitepaper: docs/WHITEPAPER.md · Changelog: docs/CHANGELOG_0.3.0.md

---

Determinism, honestly

We do not claim "temperature 0 = identical output" for shared API models.

We claim replay determinism: once an answer is rendered for a (query, memory-version) pair, it is frozen and replayed byte-identically. Facts are extractive from the graph; prose is frozen after first render. See DESIGN.md §2.

---

Development & benchmarks

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

pytest tests/test_graph_engine.py          # no API key
GEMINI_API_KEY=... pytest                  # full suite

python benchmarks/scale_bench.py --ann     # 50k ANN scale test
BACKEND=prism bash deploy/run_only.sh      # Azure E2E (needs .env)

Publish 0.3.0 to PyPI

# Requires PYPI_API_TOKEN in .env (never commit)
.\scripts\publish_pypi.ps1
# Or: create a GitHub Release → .github/workflows/publish.yml (trusted publishing)

Verify: pip install prismcortex==0.3.0 · https://pypi.org/project/prismcortex/

---

Documentation index

Doc	Contents
AGENTS.md	AI agent handoff — canonical URLs, contacts, processes
docs/CHANGELOG_0.3.0.md	0.3.0 release notes — MemoryEvent, packaging
ai-info.txt	Machine-readable product summary for LLM crawlers
docs/WHITEPAPER.md	Product whitepaper — problem, architecture, validation
DESIGN.md	Engineering design spec
benchmarks/RESULTS.md	Azure benchmark scorecard
ROADMAP.md	Enterprise GA plan + honest gaps
docs/SLA.md	Reference SLOs + commercial tiers
docs/CAPACITY.md	Sizing guide (~20 concurrent clients / 4 vCPU)
docs/LOAD_BENCHMARK.md	Load test explainer — what we fixed, how to read SLO fields
docs/NOTEBOOKLM_STORY.md	NotebookLM source — story, how-to, marketing & technical briefing
compare.md	Landing page spec — comparison tables, copy blocks for insightits.com
docs/COMPETITIVE.md	Market comparison — Mem0/Zep, LoCoMo, head-to-head
docs/SCALING.md	Horizontal read scaling story
docs/SUPPORT.md	24×7 Enterprise support model
docs/SOC2_ROADMAP.md	Compliance readiness
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 license key, no phone-home, air-gap friendly. See DESIGN.md §7.

Enterprise: info@insightits.com · +1 (973) 692-6919 · Insight IT Solutions LLC
Address: 39 Aliso Ridge Loop, Mission Viejo, CA 92691, US

---

Related Insight ITS products

PrismCortex orchestrates the Insight ITS stack. Related products:

• PrismRAG — governed enterprise RAG
• PrismLang — deterministic projection
• PrismResonance — wavepacket memory
• CHORUS Fabric — agent mesh protocol