Practical notes on AI project memory, agency team memory, benchmarks, fair product comparisons — and what actually runs in an Accord Book deployment.
The mainstream answer to a moving spec is thirty years old: a short lock, a change authority, and a living backlog. The decision those methods were making had nowhere durable to live. Coding agents make that gap operational.
Most AI memory products are an API. Accord Book is a single-organization stack you can put on your host: portal for people, MCP for agents, Postgres as truth.
TalkBase is AI project memory — one searchable brain for calls, chats, and files with RAG citations and MCP. Accord Book adds conflict detection and change-control for dev agencies. Closest search competitor; different finish line.
MemoryLake is client context memory that works across AI tools — brand voice, namespaces, skill memory. Accord Book is self-hosted project memory for dev agencies. Both say 'client context'; different jobs.
Estratos is centralized team memory for AI-assisted client work — folders, approvals, and agent connections. Accord Book is self-hosted project memory and change-control for dev agencies. Same buyer phrase, different delivery job.
Zep builds temporal context graphs for agent memory at scale. Accord Book keeps a governed project agreement for agencies. Closest technical cousin — still a different buyer.
Obsidian’s folders, links, and graph are how you look at notes. They are presentation. The durable product is collected data — and once you have that, you can present it any way you want.
Notion organizes knowledge as pages and databases. That tree is presentation. Accord Book keeps the data — then any view can be rebuilt for the team, the client, or an agent.
Mem0 is drop-in memory for agents and apps. Accord Book is a governed project record for agencies. Same word — memory — different job.
Linear tracks what to build and who owns it. Accord Book tracks what was agreed and whether new work fights that agreement. Complementary tools — not substitutes.
Letta is a self-improving agent runtime in the MemGPT lineage. Accord Book is the memory and change-control layer that governs agents you already run — not a replacement for them.
Cognee builds a company brain for agents from docs, chats, and tickets. Accord Book is change-control for agency projects — overlapping ingest, different product job.
Memory is not one product category. Here is how agent APIs, knowledge vaults, issue trackers, and Accord Book solve different jobs — and when each is the right fit.
The calculated onboarding figure on this page is $1,500 of senior time per hire. Client status calls are real hours. They are not a pile of waste the subscription deletes.
The 22 April 2026 LongMemEval slice is 20 questions: 18/20 judged correct, treatment p95 990 ms, max about 8–9 s, superseded rate 0%. Numeric gates passed. That is not a finished production-fit claim.
Graph retrieval is off the default path. LongMemEval showed zero graph contribution, and that benchmark cannot prove whether graph helps. It stays off until real pilot data shows it does.
The 6 May 2026 conflict eval is one synthetic, in-house set of 86 scenarios: 52 true positives, 0 false positives, 30 true negatives, 4 false negatives. Recall was 85.71% on the Gemini and Claude slices.
Accord Book's default retrieval stack: how vector retrieval, lexical recovery, and supersedes-aware reranking work together for current-state reconstruction in production.
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.
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.
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.
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 can draft useful documentation quickly, but publication of shared truth still needs an explicit human gate.
A practical methodology for deciding whether graph expansion is worth its operational cost in a memory system.
Useful conflict detection in software projects should surface review-worthy mismatches with evidence, not pretend to deliver perfect truth.
Why we replaced a naive contradiction prompt with deterministic candidate generation, durable findings, and a benchmarkable adjudication pipeline.
Project memory becomes useful when scattered raw inputs are normalized, structured, and kept traceable across time.
The technical and product reasons we stopped pursuing customer-hosted databases and moved to a more honest compliance story.
Long-running software projects need retrieval that handles structure, freshness, and provenance—not just semantic similarity.
AI-assisted software teams need durable memory for decisions, constraints, and provenance—not just better chat or task tracking.
A 20-question LongMemEval slice reached 18/20 judged correct and a 990 ms treatment p95. The same file shows an 8–9 s max and almost no supersedes density.
The latest Accord Book conflict-eval run found 52 true positives with zero observed false positives across 86 scenarios.