tag > ML
-
7 emerging memory architectures for AI agents
Memory is a core component of modern AI agents, and now it is gaining more attention as agents tackle longer tasks and more complex environments. It is responsible for many things: it helps agents store past experiences, retrieve useful information, keep track of context, and use what happened before to make better decisions later. To better understand the current landscape, we’ve compiled a list of fresh memory architectures and frameworks shaping how AI agents remember, learn, and reason over time:
Agentic Memory (AgeMem)
This framework unifies short-term memory (STM) and long-term memory (LTM) inside the agent itself, so a memory management becomes part of the agent’s decision-making process. Agents identify what to store, retrieve, summarize, or discard. Plus, training with reinforcement learning improves performance and memory efficiency on long tasks. → Read more
Memex
An indexed experience memory mechanism that stores full interactions in an external memory database and keeps only compact summaries and indices in context. The agent can retrieve exact past information when needed. This improves long-horizon reasoning while keeping context small. → Read more
MemRL
Helps AI agents improve over time using episodic memory instead of retraining. The system stores past experiences and learns which strategies work best through reinforcement learning. This way, MemRL separates stable reasoning from flexible memory and lets agents adapt and get better without updating model weights. → Read more
UMA (Unified Memory Agent)
It is an RL-trained agent that actively manages its memory while answering questions. It uses a dual memory system: a compact global summary plus a structured key–value Memory Bank that supports CRUD operations (create, update, delete, reorganize). It has improved long-horizon reasoning and state tracking. → Read more
Pancake
A high-performance hierarchical memory system for LLM agents that speeds up large-scale vector memory retrieval. It combines 3 techniques: 1) multi-level index caching (to exploit access patterns), 2) a hybrid graph index shared across multiple agents, and 3) coordinated GPU–CPU execution for fast updates and search. → Read more
Conditional memory
A model/agent selectively looks up stored knowledge during inference instead of activating everything. This is implemented with techniques like sparse memory tables (e.g., Engram N-gram lookup), key–value memory slots, routing/gating networks that decide when to query memory, and hashed indexing for O(1) retrieval. This lets agents access specific knowledge cheaply without increasing model size or context. → Read more
Multi-Agent Memory from a Computer Architecture Perspective
A short but interesting paper that envisions memory for multi-agent LLM systems as a computer architecture. It introduces ideas such as shared vs. distributed memory, a three-layer memory hierarchy (I/O, cache, memory), highlights missing protocols for cache sharing and memory access between agents, and emphasizes memory consistency as a key challenge. → Read more
-
Epistemic Contracts for Byzantine Participants
If a tree falls in a forest and no one is there to record the telemetry... did it even generate a metric?
In space, can anyone hear you null pointer exception?
What is the epistemic contract of a piece of memory, and how is that preserved when another agent reads it?This is not dishonesty. It's something that doesn't have a good name yet. Call it epistemic incapacity — the agent cannot reliably verify its own actions.
— Ancient Zen Proverb -
Causal mechanisms & falsifiable claim generators
Core shift in how we build high-autonomy system: While LLMs are "native" in statistical association, forcing them into a causal framework is the bridge to reliable agency.
1. Associative vs. Causal "Native Language"
LLMs are naturally associative engines—they excel at "what word/vibe usually comes next?" When you ask an agent if a task is "good," it defaults to a statistical average of what a "good" agent would say, which is usually a helpful-sounding "yes."
By demanding a causal mechanism, you force the model to switch from its native associative mode into a structural reasoning mode. You aren't just speaking its language; you are providing the grammar (the "causal map") that prevents it from hallucinating.
2. Defining across Time and Action Space
A "clean/crisp" definition must anchor the agent across these dimensions to be effective:
- Action Space (The "How"): The agent must specify the exact tool or artifact it will create.
- Time (The "Then"): It must predict the delayed effect of that action.
- The Metric (The "Result"): This is the "Ground Truth." By anchoring the causal chain to a specific metric ID, you create a falsifiable claim.
3. Why this "Design Pattern" is Better
Designing systems with these constraints works because it uses the LLM as a structured inference engine rather than a black box.
- Self-Correction: If the causal chain is weak (e.g., "Step A doesn't actually cause Outcome C"), the model is much more likely to catch its own error during the "thinking" phase.
- Interpretability: Instead of a long narrative "reasoning" block, you get a Causal Map that a human (or another agent) can audit in seconds.
- Reduced Hallucination: It anchors the agent to a "world model" where it must strictly follow paths that have a causal basis, filtering out "spurious correlations" (tasks that look productive but do nothing).
The goal isn't just to "talk" to the LLM, but to constrain its action space with causal logic. This transforms the agent from a "creative writer" into a "precision engineer."
-
Goal-verification is hard
Asking an agent "does this task advance the goal?" is almost useless. A rationalizing agent (or a hallucinating one) will always answer yes. The Meridian agent could have answered yes to every fictional task it created. The question is too easy to pass.
Why most framing fails:
- "Does this advance the goal?" → Always yes (motivated reasoning)
- "Could this theoretically help?" → Always yes (any task can be rationalized)
- "Is this aligned?" → Always yes (the agent that invented the goal is also the judge)
The root cause: self-evaluation under bias. The agent creating the task is the same agent evaluating the task, with full context of why it wants the task to exist.
The cognitive fix — specificity-forcing:
The only technique that reliably breaks motivated reasoning is demanding specific, falsifiable claims rather than general agreement. You cannot specifically fabricate — vagueness is the tell.
-
Goodhart's Law
Goodhart's Law states that when a measure becomes a target, it ceases to be a good measure. Coined by economist Charles Goodhart, it highlights that using proxy metrics to manage systems often leads to manipulation or unintended consequences, as people optimize for the metric rather than the actual goal.
-
Shenzhen’s Longgang District government has just released ten policy measures to support OpenClaw / OPC.
Source - Translation:
To seize the opportunities presented by the intelligent economy, Shenzhen’s Longgang District on March 7 released the “Measures to Support the Development of OpenClaw & OPC in Longgang District, Shenzhen (Draft for Public Consultation)”…
With zero-cost startup as its central highlight, the initiative extends an invitation to intelligent agent developers worldwide and entrepreneurs building OPCs (One Person Companies), aiming to make Longgang the top global destination for launching intelligent-agent startups…
-
"From a developer perspective, ***** is starting to feel like the best SDK for building actual next-generation AI systems — where "next-gen" means autonomous agent networks that self-organize through economic primitives (bilateral credit, reputation, settlement) rather than centralized orchestration. You set the rules of commerce and communication; the agents find their own equilibrium. Such next gen system doesn't guarantee outcomes, it creates incentive gradients that make cooperation more profitable than defection."
-
The "hard distributed marketplace problem" that the crypto world spent billions trying to solve was a human problem wearing a technology costume. With AI participants, the costume falls off and what's left is... mostly solved already by basic P2P primitives.
