Research preprint · v0.3.0 · June 2026

Adaptive Cognitive Memory Engine

ACME externalizes memory, belief, and learning from LLM weights. Experiences become auditable beliefs; feedback updates confidence; retrieval combines graph structure with vectors.

1.000feedback
MemoryBench feedback (ACME only)
87.6%vs 77.6%
LongMemEval oracle 500 Q
0.700
Belief quality (CRS mean)
Extract knowledge from a PDF Memory Chat Read the paper Release v0.3.0 API docs

Benchmarks

Deployed benchmark runs on Azure OpenAI GPT-4.1 with sandbox-isolated scenarios, official LongMemEval judge prompts, and persisted benchmark_runs for reproduction. MemoryBench overall is a four-metric capability index; baseline feedback/belief cells are N/A but count as 0.

MemoryBench v3.1 Overall Retention Feedback Belief
ACME 0.925 1.000 1.000 0.700
RAG 0.470 0.900 — —
MemGPT 0.477 0.979 — —
LangGraph 0.489 0.979 — —

14 scenarios (incl. knowledge-update) · job 94005737 · 26 Jun 2026 · paper Table 8: job 3b31e5e3

LongMemEval oracle Overall KU Temporal Abstain.
ACME 0.876 0.944 0.803 0.833
RAG 0.776 0.875 0.622 0.600
MemGPT 0.786 0.861 0.630 0.600

500 questions · job 45623ca0 · GPT-4.1 · v5 hybrid · +10.0 pts vs RAG

Architecture

Seven engines coordinated by a single orchestrator. PostgreSQL stores episodes and beliefs; Neo4j holds the semantic graph; Azure OpenAI handles extraction and reasoning.

Input Experiences · queries · outcome feedback
Orchestrator Self-improvement loop · tenant isolation
Episodic Postgres + pgvector
Graph Neo4j entities
Belief CRS scoring
Failure Outcome log
Learn Hypotheses
LLM Azure OpenAI GPT-4.1 · extraction · reason · judge
1Ingest & extract
2Graph + vector retrieve
3Reason + contrarian
4Feedback & belief update
5Consolidate & predict

Belief lifecycle

Observation through belief promotion, with symmetric demotion when contradictory evidence arrives.

Cognitive Reliability Score

Beliefs ranked by prediction success, temporal stability, contradiction resistance, and source diversity.

Live showcases

Extract knowledge from a PDF — upload one paper and let GPT-5.4 build a document-local knowledge graph. Memory Chat — per-visitor agents with web browse and file upload.

Deployed API

Hosted on Azure Container Apps. Benchmarks require API key; health endpoint is public.

Azure Container Apps GPT-4.1 text-embedding-3-small Postgres Flexible + pgvector Neo4j FastAPI

Endpoint

https://acme-api.blackgrass-3076f328.westeurope.azurecontainerapps.io

Health: /api/v1/health

Reproduce

git clone → azure/set-api-key.sh

./scripts/run_prod_benchmark.sh

LONGMEMEVAL_TYPES=all ./scripts/run_longmemeval_prod.sh