Mastra is betting that a technique called Observational Memory can make AI agents cheaper and more reliable on long, messy tasks. Instead of leaning on traditional retrieval tricks, the company says its system cuts costs by about a factor of ten while beating standard retrieval-augmented generation on long-context benchmarks. If those numbers hold up, the economics and design of everyday agents, from coding copilots to customer support bots, could change fast.
The core idea is simple but aggressive: let agents watch their own conversations, decide what matters, and store those observations as compact memories they can reuse later. That shift moves long-term context from a constant, expensive prompt to a curated knowledge base that the agent grows over time. It is a direct challenge to how most teams reach for RAG whenever they hit context limits.
What Observational Memory actually is
At its heart, Observational Memory is Mastra’s answer to long-context fatigue in agent systems. Rather than stuffing every past message into the prompt or running a heavy search over full transcripts, the agent writes down distilled “observations” about what happened, what worked, and what it learned. Those observations become a separate memory store that can be queried quickly, without dragging the full history back into every call. Mastra describes Observational Memory as its memory system for long-context agentic work, designed to sit alongside an agent and watch how the message history grows.
In Mastra’s description, the system is supported by two background agent processes, referred to as Two, that monitor the conversation stream and decide what should be turned into a durable memory. One process focuses on extracting useful facts or patterns, and the other manages how those memories are stored and surfaced later. By separating the observation step from the main agent loop, Mastra is trying to keep the core interaction fast while still building up a long-term record that does not swamp the context window.
How it differs from classic RAG
Most teams that hit context limits today fall back on retrieval-augmented generation. In RAG, the model takes a query, searches a vector index or database, and pulls back relevant documents to stuff into the prompt. That works well when the knowledge lives in static sources like manuals or wikis, but it becomes clumsy when the “knowledge” is a sprawling chat history or a multi-day workflow. The agent has to search through its own past, often with noisy embeddings and no sense of which parts were actually important.
Observational Memory flips that pattern. Instead of retrieving raw chunks, the agent retrieves its own compressed takeaways from earlier steps. A discussion thread on Observational Memory spells out the contrast directly: “Unlike RAG systems that retrieve raw documents or message chunks, Observational Memory stores distilled observations that capture what actually mattered in the interaction.” That same discussion notes that by avoiding repeated retrieval of large text blocks, the technique can push token savings into the range of 5 to 40 times on some workloads, which is where the headline 10x cost reduction figure comes from.
The benchmark results behind the 10x claim
Bold efficiency claims live or die on benchmarks, and Mastra has started to back Observational Memory with long-context tests. In public descriptions of its work, the company points to benchmarks where agents with Observational Memory were compared against agents using standard RAG setups on tasks that stretched across long histories. Those tests focused on scenarios where the agent had to remember earlier decisions, keep track of user preferences, or maintain a multi-step plan over many turns.
In those long-context benchmarks, agents built on Observational Memory reportedly outscored RAG-based agents while using far fewer tokens. A write-up shared through a long-context report describes agents that, when equipped with Observational Memory, could maintain accuracy across extended tasks while cutting total token usage by roughly an order of magnitude. The same reporting ties those savings to the way Observational Memory prunes away redundant context, so the model no longer has to reread the same backstory every time it answers a follow-up question.
Why the cost drop matters for real products
For teams building production agents, a 10x reduction in token usage is not a nice-to-have, it changes what is economically viable. Long-running support agents that follow a customer through multiple tickets, coding copilots that track a project over weeks, or workflow bots that manage complex onboarding flows all face the same problem: context gets long, and prompts get expensive. If an approach like Observational Memory can keep the agent grounded in what happened earlier while trimming most of the prompt, whole classes of “too expensive” ideas start to look feasible.
The impact is particularly clear in areas like legal intake tools or health coaching apps, where each user’s history matters but margins are thin. Instead of paying to resend a full transcript or case history, an Observational Memory-style system can retrieve a small set of past observations, such as “user prefers written summaries over calls” or “patient has already tried medication A and reported side effects.” Mastra’s own Observational Memory documentation frames this as a way to keep agents “aware” of long histories without dragging the entire message log into each call, which is exactly the kind of tradeoff product teams need.
Limits, open questions, and what comes next
None of this means RAG goes away. Retrieval over external sources like Confluence pages or GitHub repositories still solves a different problem than remembering how a specific user conversation unfolded. In practice, many systems are likely to combine both: Observational Memory for the agent’s own experience, and traditional retrieval for outside documents. The tricky part will be orchestration, deciding when to lean on observations, when to search, and how to keep those two views of “what the agent knows” consistent.
There are also open questions about failure modes. If an observation is wrong or biased, the agent might keep reusing that flawed memory. Designers will need tools to inspect, edit, and sometimes delete observations, much like they manage logs or analytics events today. Mastra’s own memory system hints at this by emphasizing how the background processes watch the message history as it grows, which implies there is room for smarter filters, better summarization models, and governance. If those pieces mature, Observational Memory could move from a clever optimization to a standard pattern for long-lived AI agents.
