π0.5 (Physical Intelligence)

https://www.pi.website/blog/pi05

Extends π0 with MEM layer giving the policy short-term (within-task) + long-term (cross-task, >10-min horizon) memory. Co-trained on robot teleoperation + human video + text. Hierarchical inference: VLA predicts subtasks as text tokens, then executes as action chunks.

At a glance

Type
Multi-Scale Embodied Memory (short + long episodic)
Tier
T1
Created
2024 (Physical Intelligence formally founded 2024; informal work began Q3 2023; Chelsea Finn Sergey Levine Karol Ha…
Latest release
not applicable — not OSS
License
not applicable — not OSS
GitHub
not applicable — no GitHub repo
Pricing
searched not found
Funding
$400M total $2.0B val Series A · 2024-11

Taxonomy

storage
vector
retrieval
attention
persistence
long-term
update
append-only
unit
episode
governance
opaque
conflict
append

When to use

Optimised for: real-time perception + spatial reasoning + multi-modal sensor fusion

Anti-fit: not for non-embodied / non-physical use cases

Pros & cons

Pros

Generalist robot policy with strong simulation-to-real transfer; memory is implicit in the policy network rather than explicit external store.

Cons

Implicit memory is harder to inspect or edit; embodied agent only — not generalizable to LLM agents.

Claims & capabilities

Open-world generalisation to unseen household environments. Cleans an entirely new kitchen with no environment-specific training.

Technical surface

API surface
searched not found
Backend storage
searched not found
Deployment
Cloud inference + on-robot deployment (websocket streaming)
Embedding model
searched not found
Multi-tenancy
searched not found
MCP
no MCP support advertised — vertical product, no MCP server / client integration documented
A2A
no A2A protocol support advertised — vertical product, no A2A integration documented
OpenTelemetry
no OpenTelemetry integration advertised — vendor logs/observability not publicly documented

Similar systems

Other vertical / domain-specific ai memory in the catalog, ranked by inbound references.

  • NVIDIA ReMEmbR T3

    Builds long-horizon memory by captioning video segments with VILA, storing captions with timestamps + 3D position coordinates in MilvusDB. At query time, LLM iterates retrieval across text, time, and position modalities. Deployed on Nova Carter robot (Jetson Orin).

  • Abridge T1

    Clinician-assist ambient documentation. Source mapping: every AI-generated summary element traced back to the source utterance. Audit-and-trust layer over episodic memory. Built on proprietary 1.5M+ medical-encounter dataset.

  • ASAPP GenerativeAgent T1

    Treats memory as first-class architecture. Captures the digital footprint of every interaction; retrieves preference and history at engagement time. Public example: airline knowing a frequent flyer wants aisle seats with her son — preference-aware, not just history-lookup.

  • BenevolentAI T1

    Target identification / drug repurposing / mechanism tracing. 85+ data sources, petabyte-scale, rebuilt every few weeks. Wet-lab results re-enter the graph and shift downstream predictions — institutional experimental memory closing a feedback loop.

  • Causaly T1

    Drug discovery / target identification / causal mechanism tracing. The graph is the memory: 7 years of curated biomedical cause-effect relationships compounding with each new ingestion. Scientific RAG retrieves from a versioned causal substrate.

  • Character.ai T1

    Chat Memories (user-defined facts), auto-memories for c.ai+ subscribers, pinned memories, in-context retention. PipSqueak 2 model (April 2026) reduces in-conversation drift. Memory Visualization meter forthcoming.

Row last verified 2026-05-14. Catalog data is CC-BY-4.0 — see how to read this.