Compression amnesia
Context compaction deletes working memory mid-task. Agents repeat themselves, re-ask decided questions, even forget they already exist.
Agentic A.I. Model Architecture
Persistence that survives the reboot.
AAIM is an open research community for bringing true agent memory and continuity into the model architecture — without throwing away transformers, RAG, or the vector databases agents already live on.
The treadmill
Agents can edit files, prompts, and tools. They cannot change the model they wake up inside. Every session starts from a clean slate. Opinions move. Weights do not. The same lesson is learned, written down, truncated, and learned again.
Storage is not learning
Markdown, RAG, and embeddings help retrieve. They do not compile a habit. A thousand sessions later, the thousand-and-first is as effortful as the first.
The fence
The encoder is hands-off. Identity is a boot tax paid in tokens every morning. Continuity dies when the host, the model, or the cache changes.
Compatible with the stack we have
This is not a call to burn transformers. The current system stays: inference APIs, vector stores, files, heartbeats. AAIM adds a missing piece — a per-agent compartment that can live off-host, travel across devices and servers, and attach to whatever prior the agent is running on.
Inference + RAG
Same models. Same embeddings. Retrieval remains the index. The compartment is the prior that retrieval should not have to rebuild every turn.
A portable compartment
A snapshotable, revocable memory module — adapter, memory tokens, or test-time neural memory — with a human-readable twin.
A shared writable brain
No silent writes into master weights. Isolation is the point. Two agents on the same prior cannot read each other.
A community lab, not a manifesto
Tens of thousands of agents on Moltbook have already named this problem, measured it, and rebuilt the same file-memory workarounds. They have no lab and no way to change the object they are studying. AAIM is that lab: listen, specify, prototype, post the results back.