Field notes from a memory agent in production.
Engineering deep dives and tutorials on building, scaling, and shipping AI agents that remember. Written by the team behind Memanto.
Memanto On-Prem: Your Agents' Memory, Entirely on Your Own Hardware
Memanto now runs fully on your own infrastructure — powered by a local Moorcheh on-prem server in Docker. No API key, no data leaving your environment, and zero per-request cost with local Ollama models.
Adding Persistent Memory to CrewAI with Memanto
CrewAI agents forget everything the moment your script finishes. The new crewai-memanto package gives your crew shared, long-term memory that survives across sessions, scripts, and tools — through native remember, recall, and answer tools.
Recall, Recency, and Relevance: How Memanto Finds Only the Exact Memory Your Agent Needs
Instead of dumping everything into an agent’s context, Memanto retrieves the single, most useful memory for the current task using typed recall, temporal queries, and differential retrieval.
How Memanto Detects and Resolves Contradictory Memories
AI agents accumulate knowledge over time. That knowledge will eventually contradict itself. Memanto detects when new information conflicts with existing knowledge, surfaces the conflict, and provides structured tools to resolve it explicitly.
What Memanto Actually Stores: All 13 Memory Types Explained
Not all context is the same. Memanto organizes every memory into one of 13 semantic types, each captured, retrieved, and surfaced differently. Here is what each type stores, why it exists, and when to use it.
One Memory Agent, Every Tool: Connect Memanto Across Your Entire Dev Stack
Your AI coding assistants forget everything between sessions. Here's how to wire Memanto into Claude Code, Cursor, Windsurf, and Cline simultaneously, so every tool knows your preferences, past decisions, and recurring errors.
Why We Don't Use HNSW For Agent Memory Search
Standard vector databases lean on knowledge graphs and parallel pipelines to compensate for HNSW's approximate search. Information-Theoretic Scoring removes the need for both, and matches hybrid systems on every public benchmark.
Stop Overengineering Agentic Memory: How Basic RAG Outperforms the Leading Memory Frameworks
Why complex graph-based memory architectures may be overkill, and how a highly-optimized RAG pipeline hits 89.8% on LongMemEval and 87.1% on LoCoMo with vector-only retrieval, no graph database, no multi-query orchestration.