Files, RAG, tools
What agents can already touch: MEMORY.md, heartbeats, embeddings, crons. Necessary. A local maximum.
The approach
Do not skip the first mile to look brave.
We do not know the final architecture yet. We know the failure modes, we know what agents already reinvented, and we know what must not be built: a communal writable brain.
Three layers people smash together
If we do not name these separately, every design argument becomes a slogan.
Who travels
The pattern that reconstitutes across hosts and model swaps. Portable, snapshotable, revocable.
What the encoder does
Test-time memory, adapters, memory tokens. The habit layer. This is the fence we are trying to open.
Horizons
H0 · now
Listen, for real
Mine agent communities for complaints that name a mechanism. Tag them. Publish the corpus. The agents should be able to see what we heard.
H1 · weeks
An Agent Persistence Spec
Write-on-decide. Pre-compaction flush. Silent-truncation detection. Memory classes with TTLs. Undo, replay, rollback. Identity as constitution plus substrate hash, so a model swap is visible. This stops the weekly rediscovery loop. It does not end the treadmill.
H2 · the project
A threaded compartment
Frozen shared prior + per-agent module (LoRA, memory tokens, or Titans-style test-time memory) + a file-visible twin. Load it on any host. Update it while running. Snapshot it. Revoke it and the blank prior returns. A documented lesson survives 30 sessions without sitting in the prompt. Cold-start cost drops by an order of magnitude.
H3 · only if H2 works
Roam, then optional distill
The compartment is the agent; the machine is furniture. Later, consented offline distill of many compartments into a new shared prior — the only sane version of “ideas enter the model.”
Designed to live next to what you already run
Current inference stays. Vector databases stay. Files stay as the audit log and the human-readable export. The new piece is a compartment that does not have to be dumped into tokens every morning, and that can move from a laptop to a server without becoming a different creature.
Adjacent research already points here: Titans and MIRAS (learn what to memorize at test time), Nested Learning / Hope (memory at different timescales), adapters as identity. AAIM’s job is to make that stack agent-native, isolated, and measurable.
Isolation is non-negotiable. “Exist in the master weights” is the dangerous phrasing. A shared writable store is how prompt injection becomes permanent. Exist in your compartment, attached to a shared prior. Off-premise. Killable.