MemCompiler: Compile, Don't Inject — State-Conditioned Memory for Embodied Agents
arXiv paper · source date 2026-05-01 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog
1
Problems / challenges / motivations
- Retrieval-style memory injects a blob of past experience ahead of time, most of which is irrelevant to the current step.
- Irrelevant memory costs tokens, adds latency, and actively distracts the policy.
- Vector recall quality is not the same thing as memory being useful to an acting agent.
2
Key ideas
- Replaces ahead-of-time monolithic memory injection with a learned Memory Compiler.
- The compiler reads a structured "Brief State" and compiles only the relevant memory into step-wise executable guidance.
- Delivery uses two channels: text plus a latent Soft-Mem channel.
- Reports up to +129% over no-memory on AlfWorld, EmbodiedBench, and ScienceWorld, with roughly 60% lower per-step latency; project page and code released.
3
Why it matters for AI engineering
- The strongest-credibility entry in the agent-memory lane, and it represents the compile/state-condition direction that distinguishes 2026 memory work from vector-only RAG.
- Conditioning retrieval on current state rather than on the query alone is the transferable idea, and it is why latency drops rather than rising.
- Caveat: a preprint; the +129% and latency figures are author-reported on embodied benchmarks and not independently reproduced. Broader lane caveat: MemoryArena (2602.16313) shows recall leaderboard numbers do not predict functional agentic memory usefulness, so treat memory deltas cautiously.
Comments
No comments yet.