BLOG
AI Memory & Agent Engineering — Field Notes from Maximem
Research, product releases, and engineering notes on AI memory, agent context, and production AI infrastructure.
AI Agent Memory
See all 66 →AI agent memory and context engineering guides
Practical guides, engineering patterns and mental models for building agents that remember, reason and act.

How to Migrate Agent Memory: Import Existing User History Into a New Memory Layer Without a Gap
How to migrate agent memory to a new memory layer without users noticing: sort history into raw chats, extracted facts and profile records, keep every original timestamp, import oldest first per user and one session per call, make the import idempotent on source ids, filter secrets at the door, then cut over with dual writes, a background backfill, answer comparison on real questions, staged reads and a read-only rollback store. With the export paths Mem0, Zep, Letta and Cognee document as of 18 November 2026, and how to re-index when the embedding model changes.

Agent Memory Retrieval Ranking: Relevance Is Not Usefulness, and How to Choose What Goes Into a Small Context Budget
How to rank agent memories when the context budget is small: use relevance to collect candidates and rank the final set on what each memory changes in the next decision. In MemCredit's ALFWorld study (arXiv 2611.15600, authors' report) 86.9% of memory interventions changed nothing and relevance correlated at -0.011 with usefulness; ranking by measured usefulness reached 86.57% success at a 500-token budget against 65.67% for BM25. What to rank on besides similarity, how to measure usefulness on your own traces with ablation, budget packing, the limits, and a tested Python build.

Agent Memory Versioning: How to Roll Back a Bad Memory Write, Restore the Index With It, and Find the Answers That Relied on a Changed Fact
Agent memory versioning keeps every change to what an agent remembers as an immutable revision, so a bad write can be inspected and reverted instead of silently overwriting the truth. What to record per revision, why a rollback has to restore the retrieval index and derived facts along with the stored memory, how Anthropic, Google Vertex AI, Mem0, Letta and Hindsight handle history and rollback as of 13 November 2026, how versioning interacts with GDPR erasure, and how to run provenance backwards to find the answers already sent that relied on a fact that has since changed, plus why a self-improving agent's edit history belongs outside the model. With a tested Python build.
Context Engineering
See all 26 →
Async Subagents: What the Orchestrator Keeps in Context, What Workers Hand Back, and What Should Outlive the Session
Async subagents let an orchestrator stay reachable while background workers run long tasks. What the orchestrator must keep about in-flight work so compaction cannot erase it, what a worker should hand back, when a follow-up should interrupt or enqueue, how results reach an orchestrator that is busy, idle or waiting on an approval, and which decisions and preferences should outlive the session, with LangChain Deep Agents, Claude Code and OpenAI behaviour as of 6 November 2026.

Context Compaction vs Retrieval: When Summarising a Long Agent Session Is Cheap, When Exact Lines Win, and What Neither Does Across Sessions
Compacting a 400,000-token Claude Opus 5.5 session costs about $0.08 of input with a warm prompt cache and $1.60 to $2.00 with a cold one, at list prices as of 11 November 2026. Mainplane's measurement, NVIDIA's evidence that agents can compact after every action, what that assumes about state kept outside the window, Claude Code's and OpenAI's compaction docs, and the cross-session problem neither approach solves.

LLM Context Window Sizes in 2026: Every Model's Limit, the Long-Context Price Cliffs, GPT-4o's 128K, and What Happens When a Chat Runs Out
Every major LLM's context window and max output as of 28 October 2026, from GPT-6's 1,050,000 tokens to Grok 4.7's 500,000, the long-context price cliffs at 100K (Claude Haiku 5.5), 200K (Gemini 3.1 Pro, Grok 4.7) and 272K (OpenAI), GPT-4o's 128,000-token window and 16,384-token output limit, and what each API and chat app does when you run out.
RAG & Retrieval
See all 13 →MCP & Agent Protocols
See all 5 →Integrate memory across your stack.
Guides and examples for using Maximem with MCP, SDKs and your existing tools.

What is WebMCP? When and How to Use WebMCP in a Browser Agent?
The decision is per step, not per site: a page can expose a clean search tool and leave its account settings as ordinary DOM controls, so an agent that picks one mode per domain gets the worst of both.

MCP 2026-07-28: 20 Breaking Changes and the Errors They Cause
MCP 2026-07-28 removed sessions, the initialize handshake and server-initiated requests, and Claude Code now speaks it by default. The full diff, every error you will hit with the fix for each (including 401, invalid_grant and -32000), how to fail so an agent can recover, the DCR to CIMD migration, changing tools without breaking running agents, SDK status on 5 November 2026, and the HTTP+SSE deadline two official sources still disagree about.

MCP Servers Explained: What They Are and How AI Agents Use Them
An MCP server is a program that exposes tools, resources and prompts to an AI application over the Model Context Protocol. Updated 6 November 2026 for the current spec, 2026-07-28: stateless requests, stdio and Streamable HTTP, OAuth 2.1 for remote servers, and a practical guide to choosing between tools, resources and prompts, including where agent memory fits.
Agent Frameworks
See all 17 →Agent Evals & Observability
See all 12 →
How Do You Test Agent Memory Before You Ship? A Test Plan You Can Run in a Week
A one-week test plan for agent memory: script multi-session histories with a known answer key, then grade the store, retrieval and outcome separately across recall, continuity, contradiction, staleness, deletion, cross-tenant leakage and hallucinated recall, with a metric and a recommended pass bar for each.

AI Agent Evals: The Judge Is The Part Nobody Measures
An LLM judge scoring the same agent output three times will often hand back three different answers, and almost no guide to agent evaluation asks whether the judge agrees with itself or what it costs to run on everything rather than a sample.

How to Evaluate AI Agents (and Test a Model Swap Before You Switch)
How to evaluate an AI agent: start with 20 to 50 tasks drawn from real failures, grade the outcome and the transcript, run several trials per task, gate every change in CI, and check the grader before you trust it. Updated 21 November 2026 with what evals cost and how to cut the bill, alongside a seven-step method for testing a cheaper model inside your own agent before you switch.
Claude Code & Coding Agents
See all 11 →Voice Agents
See all 4 →
I Spoke to 500+ Voice AI Builders in India Over 3 Months. Here Is What I Found.
Field notes from 500+ Voice AI builders in India. An exploration of outbound dominance, the "too good" TTS problem, and why memory is the final infrastructure hurdle for production agents.

Voice Agent Stack: The Right Tools for Production Voice AI in 2026
Build production-ready voice agents with the right stack. Compare MCPs, frameworks (CrewAI, AutoGen, Swarms), APIs (Deepgram, ElevenLabs, Vapi), and learn cost-effective patterns for voice AI in 2026.
LLM Cost & Production
See all 27 →Research & Product Updates
See all 12 →
Synap Is #1 on LongMemEval: 95.6% on All 500 Questions
Maximem Synap is #1 on LongMemEval: 95.6% on all 500 questions with Gemini 3.8 Flash, at about 7.4K tokens per question. Every answer and verdict is public.

Maximem Synap Updates: Higher Scores, 17 Integrations, and a Live Playground
Synap updates: 95.6% LongMemEval, #1 on the benchmark (up from 92% and 90.2%), 93.2% LOCOMO, 17 framework integrations, a browser playground, public pricing, and a free accuracy eval on your own agent.

How Synap Works Under the Hood
We launched Maximem Synap today. Here's a peek into how it is built.
Beyond our own writing, see our curated directory of engineering blogs.
Build agents that remember.
Add persistent memory and context to your AI agents with Maximem Synap.














