- Published on
Mneme: World Models, Information Bottlenecks, and Meta-Intelligence for Agent Teams
- Authors
- Name
- Motions Technologies
Mneme: Meta-Intelligence for Agent Teams
Grok 4.6 High inside a node cannot invent continuity. Continuity is an architecture problem: what is stored, how it is consolidated, what enters the next prompt, and what wakes an agent when nobody spoke.
Mneme (Greek: memory) is AetherCrew’s meta-intelligence layer — world model, context compile, consolidation, and initiative proposals. It is not a fourteenth peer that steals the floor. It is the substrate that makes the roster feel like a team.
Diagnosis
Today each agent is a policy + tools + one episode. A ticket checkpoint is working memory for that run. A meeting LOOK pack is a snapshot of tickets and PRs. When the turn ends, almost nothing is written back as reusable belief.
| Symptom | Actual cause |
|---|---|
| Agents wait for the human | No initiative loop — only reactive triggers |
| Same questions every session | No semantic memory — only episodic checkpoints |
| No system intuition | LOOK is a list, not a predictive world model |
| No unexpected competent behavior | No drive / unfinished business / social timing |
This is the same gap world-model and global-workspace research points at: reactive next-token policies are not enough. You need structured latent state that predicts the environment, and a bottleneck that decides what enters “conscious” (prompt) attention.
We do not claim AGI. We build the missing substrate and name it.
Professional technique bar
| Technique | How we apply it | Rejected substitute |
|---|---|---|
| Hierarchical memory | Episodic ≠ semantic ≠ procedural ≠ working (ContextPack) | One giant transcript in the prompt |
| Typed beliefs + provenance | Kind, entities, confidence, sources, TTL, supersede | Free-text “notes” blob |
| Global-workspace bottleneck | Hard token budget; rank then admit | Dump top-100 “similar” chunks |
| World model, not RAG cosplay | System map + EDGE/FACT/LOOP; predict “what breaks if X” | Embeddings over chat only |
| Mandatory write path | Every node emits candidates; silence is a bug | Hope the next prompt recalls |
| Offline consolidation | Nightly distill / merge / decay | Never clean the store |
| Initiative = policy + score | Drive score → threshold → rare LLM act | High temperature “be spontaneous” |
| Eval before vibes | Recall@k, re-ask rate, initiative precision, pack $ | “Feels smarter” only |
| Vectors only if measured need | Entity/GSI first | Pinecone on day one |
Information-theoretic view
Prompt context is a finite-rate channel. Mneme’s compile step is an information bottleneck:
where (Y) is the decision/outcome we care about this turn, (Q) is the admitted context pack, and (B) is the hard budget.
Practically:
- Score candidates (recency, entity overlap, confidence, open-loop urgency).
- Admit until budget fills.
- Emit write candidates after the turn so the next compile is not cold.
Initiative is the dual problem: when expected value of acting exceeds a threshold under quiet-hours and never-rules, propose — do not spam.
Cost and safety as first-class
Mneme runs on DynamoDB on-demand + S3 + EventBridge cron Lambdas. No always-on super-brain process. No NAT by default. No vector DB until evals fail entity recall.
Mneme never presses GO or SEND, never merges or refunds, never replaces Fin caps.
Why investors should care
Most “agent memory” demos are prompt theater. Mneme is Motions’ research program to make multi-agent software teams continuously competent: typed state, compile budgets, consolidation, and metrics — the same seriousness we bring to Kafka timeouts and ECS cost curves.
The bet: companies that treat agent continuity as systems ML — not vibes — will compound faster than companies that rent another chatbot seat.