Memory in Robotics: What Transfers to AI Agents, and What Doesn’t

What robotics has actually solved

Four memory problems robotics has real answers to:

1. Spatial persistence (SLAM)

Robots map an environment and localize within it across sessions. The map survives power cycles. The equivalent in AI-agent terms: your agent should know the "map" of an organization’s entities, relationships, and state after a power cycle - which is exactly what a context graph does.

2. Episodic memory

Endel Tulving’s framework for memory distinguishes episodic (specific events with time and place) from semantic (general knowledge). Robotics research since the 1990s has built episodic memory systems that store events with spatial and temporal context - "I saw the red ball at 3pm near the kitchen" vs "red balls exist." The AI-agent equivalent: distinguishing "Jordan replied at 3:15pm on Tuesday" from "Jordan is a customer."

3. Continual learning

How robots acquire new skills without forgetting old ones ("catastrophic forgetting"). Techniques: elastic weight consolidation, experience replay, progressive networks. The AI-agent equivalent: how self-evolving agents (Hermes, Letta) update skills without overwriting good procedures.

4. Hierarchical representation

Robots don’t store every pixel of every observation. They abstract - floor, room, building, neighborhood - and retrieve at the right level. AI-agent context systems increasingly do the same: Short/Medium/Long retrieval tiers, progressive disclosure of skills.

What doesn’t transfer

• Pure SLAM. SLAM solves the problem of "where am I?" AI agents don’t have a physical location. Trying to port SLAM into a chatbot is a category error.
• Sensor fusion. Robotics fuses LIDAR, IMU, cameras. AI agents don’t have these. The analogous problem - fusing signals from Gmail + Calendar + Slack + HubSpot - looks similar but is not the same computational problem.
• Motion planning. A* search, RRT, optimization-based controllers. These solve "how do I move from state A to state B?" AI agent "planning" is symbolic, not kinematic. The algorithms don’t transfer.
• Embodiment-specific cognition. The whole "4E cognition" thesis (embodied, embedded, extended, enacted) assumes the cognitive system is physically situated. For software agents, this is philosophy, not engineering.

The concrete patterns to steal

Three robotics patterns worth porting directly:

Pattern 1 - Place cells and map-building

Hippocampal place cells (O’Keefe’s Nobel-winning work) inspired robotics to build grid-based maps that persist across sessions. AI agents can do the analogous thing: entity-based "maps" where each entity (person, deal, project) has a persistent representation, and the agent builds a richer map over time. This is what every serious memory layer is converging on.

Pattern 2 - Experience replay

Robots store raw sensor data plus outcomes, and periodically replay them to improve. AI agents can store raw signal data plus outcomes (what was the recommendation, did it get acted on, what happened after), and replay weekly to extract precedents. Genios’s precedent-extraction job runs this pattern directly.

Pattern 3 - Active perception

Robots don’t passively observe - they move to get better data. AI agents can "move" too: ask follow-up questions, call specific tools, request clarification. Good harnesses build this into the agent loop.

Where robotics AI and agent AI are converging

Two areas where the problems are structurally the same:

1. 01 Continual learning in production. Both fields have agents that run for months and accumulate experience. Both face catastrophic forgetting when the system is retrained or updated. Solutions are converging on similar techniques (parameter freezing, experience replay, hierarchical consolidation).
2. 02 Temporal reasoning about changing state. A robot needs to know "the box was on the table at 3pm, but it’s gone now." A CRM agent needs to know "Jordan was CTO at BrightPath, but they moved to a new role two months ago." Same temporal-validity problem, different surface form. Zep / Graphiti’s temporal knowledge graph is the robotics-adjacent answer.