What Accord Book Actually Costs — And What It Saves
Flat annual pricing sounds simple. Here is the math behind why it is cheaper than the status quo even before you count the subscription.
Technical Blog
Practical notes on AI memory systems, benchmarks, and deployment decisions.
Accord Book Retrieval Benchmarks: What the April 2026 Run Showed
Accord Book's April 2026 LongMemEval-style evaluation cleared all production-fit thresholds: 990ms retrieval p95, 90% judged accuracy, and 0 API errors — validated end-to-end through the full pipeline.
Graph Off the Hot Path: What the Benchmarks Showed and Where Graph Now Fits
After benchmarking graph contribution across every run, we retired graph from the default retrieval lane and repositioned it as a derived intelligence layer for provenance, change impact, and project reasoning.
Accord Book Conflict Detection: 100% Precision, 92.86% Recall on May 2026 Benchmark
On the May 6, 2026 conflict-eval run, Accord Book recorded 52 true positives, 0 observed false positives, 30 true negatives, and 4 false negatives across 86 scenarios.
Accord Book's Retrieval Stack in Production: Three Lanes, No Graph on the Hot Path
Accord Book's default retrieval stack: how vector retrieval, lexical recovery, and supersedes-aware reranking work together for current-state reconstruction in production.
Accord Book's Docs Pipeline: arc42, ADRs, and PR-as-Publication
A practical look at Accord Book's Git-backed documentation pipeline: read-only repo analysis, bounded evidence packs, staged synthesis, and PR-based publication for `.q_context` docs.
Why Production Retrieval Needs More Than Vector Search
A technical look at why production retrieval for long-running project systems depends on multiple search lanes, provenance, freshness, and current-state reasoning rather than embeddings alone.
How to Generate Architecture Docs from Real Code Without Turning Your AI Into a Build Agent
A technical look at how to generate architecture documentation from real repositories using read-only analysis, bounded evidence, and human-gated publication instead of code execution.
Designing Conflict Detection That Earns Trust
A technical look at why conflict detection in project systems needs structured candidate generation, evidence-backed adjudication, and a clean separation between findings and alerts.
AI Proposes, Humans Publish
AI can draft useful documentation quickly, but publication of shared truth still needs an explicit human gate.
How We Benchmark Graph vs. Vector-Only Retrieval Before Pricing the Product
A practical methodology for deciding whether graph expansion is worth its operational cost in a memory system.
Designing Conflict Detection as a Risk Signal
Useful conflict detection in software projects should surface review-worthy mismatches with evidence, not pretend to deliver perfect truth.
From Single-Hop LLMs to Two-Stage Conflict Detection
Why we replaced a naive contradiction prompt with deterministic candidate generation, durable findings, and a benchmarkable adjudication pipeline.
From Scattered Inputs to Usable Project Memory
Project memory becomes useful when scattered raw inputs are normalized, structured, and kept traceable across time.
Why We Abandoned BYODB and Chose Portability Over Isolation
The technical and product reasons we stopped pursuing customer-hosted databases and moved to a more honest compliance story.
Why Hybrid Retrieval Beats Plain Vector Search
Long-running software projects need retrieval that handles structure, freshness, and provenance—not just semantic similarity.
Why AI Teams Need a Memory Layer
AI-assisted software teams need durable memory for decisions, constraints, and provenance—not just better chat or task tracking.
LongMemEval notes: Accord Book retrieval is now above the production-fit line
Our latest memory-eval run reached 90% post-curation LLM-judged correctness while staying under the retrieval latency threshold.
Conflict detector benchmark: high precision, with honest recall caveats
The latest Accord Book conflict-eval run found 52 true positives with zero observed false positives across 86 scenarios.